AI & Machine Learning

The Blueprint: Orchestrating 5,000+ AI assets for a World Cup TV spot

Fri, 06 Nov 2026 17:00:00 +0000

Welcome back to The Blueprint, a regular feature where we highlight how Google Cloud customers are tackling unique and common challenges across industries using the latest AI and cloud technologies. We hope to inspire others looking to innovate in their work.

AI is rapidly transforming video production at virtually every step of the process, from ideation and storyboarding to filming, editing, and release. At Mito, we’re striving to democratize AI-powered video creation, enabling everyone from solo artists to large enterprises to get the most out of the best creative technologies available.

The Challenge

When we founded Mito, our goal was to eliminate the complex technical friction that stops creators from telling long-form, narrative stories with generative AI.

However, transitioning from a single five-second clip to a cohesive campaign — such as a 120-second national TV spot for Movistar Plus+, a part of Telefónica, to broadcast across Spain during the World Cup this summer. This effort revealed massive infrastructure and workflow bottlenecks:

Asset & workflow overhead: Managing local media libraries across individual workstations was unscalable. For Movistar alone, our production required generating a massive "iceberg" of over 5000 assets (1000 of those videos) generated in around 65 iteration rounds.
Zero real-time collaboration: Traditional feedback loops required manual downloads, Google Drive uploads, and fragmented communication across chat apps and email.
Rendering latency: Standard AI video rendering takes four to five minutes per short clip. Running these processes synchronously froze our user interface and stalled creative velocity.
Model disparity & consistency: Every generative model requires distinct prompting logic. Maintaining character persistence, specific lighting, and camera angles across dozens of shots — while preserving the likeness of real public figures like chef Ferran Adrià or journalist Juanma Castaño — was nearly impossible without central orchestration.
Capacity constraints: Operating on public pay-as-you-go model APIs left us vulnerable to rate limits and regional throughput throttling during peak traffic.

The Solution

To deliver an infinite, real-time creative canvas for filmmakers, we designed a serverless, asynchronous architecture built on Google Cloud.

We run our API, AI Director, and background workers as Java services on Cloud Run, which scales with request volume.

In order to address message routing challenges, we decoupled job execution using Google Cloud Pub/Sub as our central event bus. Long-running video generations are offloaded to background Cloud Run workers, allowing creators to keep working uninterrupted on the UI canvas.

To minimize latency, media assets are stored in Google Cloud Storage and delivered via Cloud CDN. Expiring key signatures keep pre-release client assets strictly private. Keeping our storage alongside the models we call through Vertex AI reduces network transfer delays.

An isolated database architecture helps with the heavy business intelligence and analytics workloads while mitigating potential negative impacts on production performance through Cloud SQL and PostgreSQL. We rely heavily on foundational models like Omni, Veo and Nano Banana, which we access through Vertex AI alongside our custom system prompts.

Finally, our structured canvas frontend enforces a "Rule of 6" (limiting prompt variables across World, Camera, Light, Texture, Actions, and Character) and prioritizes still-frame generation before video rendering to maximize cost efficiency.

The Architecture

The Outcome

By pairing our creative orchestration platform with Google Cloud’s infrastructure, we successfully delivered a landmark national broadcast campaign that the production team describes as 95% AI-generated:

Prime-time scale: The 120-second Movistar campaign for the Spanish National Team aired across every major television channel in Spain, with 1,441 assets live on a single shared Canvas near launch.
AI Director copilot: Our conversational agent translates natural language instructions into precise model parameters while maintaining context, lighting, and visual style across every scene.
Seamless collaboration: Distributed teams and brand clients can work on the same canvas in real time. Because new generations are saved as separate assets instead of overwriting earlier ones, stakeholders can duplicate projects to explore parallel creative ideas.
Continuous dogfooding: Through our internal studio, Mito Films, we put our tools to the test on high-stakes client projects — whether that's global television spots or keynotes for top robotics startups — ensuring direct feedback flows back into engineering.
Future-proof foundation: Our architecture is primed for rendering times to collapse from minutes to seconds, positioning Mito to handle upcoming world models, deeper memory states, and personalized content delivery.

The Details

Building an enterprise-grade platform for AI storytelling required solving several backend and creative challenges. This includes our creation of a virtual queuing system that mitigates third-party API rate limiting during usage spikes for smooth request concurrency, with bounded retries and a clear outcome for every job.

Because creative output cannot be fully validated with standard unit tests, we built an automated evaluation harness powered by Gemini. Acting as an independent judge, it evaluates AI Director outputs against original user intents across an automated test suite, which we run alongside standard tests for high-impact Director changes.

We knew early on that we would need multi-model intelligence routing. As such, we built the AI Director to dynamically coordinate tasks across multiple specialized models — including Sonnet, Gemini, and ElevenLabs for audio — treating the spatial canvas like a live code repository to infer adjacent contextual cues.

We also needed to embrace an “imperfect aesthetic” to arrive at a more authentic outcome. In our Movistar spot, the story follows Tommy, a 17-year-old student making an AI video project. At the climax, the AI misunderstands a prompt, making coach Luis de la Fuente say "In the same milkshake" while holding a drink. Tommy edits the prompt live to "In the same heartbeat," and the video recalibrates instantly.

By incorporating surreal visual elements — like stadium stands built into Galician cliffs or a neighborhood bar covered in pitch grass — we showcased how generative AI can elevate human storytelling rather than obscure it.

Bringing conversational analytics to Gemini with Agentic Data Cloud connectors

Thu, 05 Nov 2026 23:00:00 +0000

Data Cloud connectors in Gemini for business, now in preview, give business users and analysts a direct way to query live enterprise data in natural language. Across most organizations, data questions multiply faster than analysts can answer them: Projects stall in ticket queues, leaders fall back on guesswork, and unmanaged spreadsheets spread across teams. Early generative AI tools promised conversational access to data, but they often struggled to generate accurate SQL across complex schemas or to respect governance rules. Data Cloud connectors let you query governed data directly from your workspace while enforcing row- and column-level security and respecting end-user credentials.

A new way to query enterprise data

Gemini brings conversational analytics directly into your day-to-day workflow through two paths: Conversational Analytics agents, which let you create data agents that include context and instructions to effectively answer questions for specific use cases, and Data Cloud connectors, allowing you to query live data directly with Gemini. End users can pull data directly from BigQuery (preview), operational databases like Spanner, AlloyDB, and Cloud SQL (preview), and governed data from Looker (preview). From there, you can move your findings directly into everyday tools, such as turning query results into a Google Doc, an email draft in Gmail, or a deck in Google Slides.

Ask natural-language questions across enterprise data using Data Cloud connectors.

How Data Cloud connectors work

Data Cloud connectors work with Knowledge Catalog to locate relevant data, interpret your question, and execute permission-scoped queries under your identity.

Knowledge Catalog interprets the user prompt and identifies the right data source for the query.

When you ask a question in Gemini, the system coordinates discovery, semantic interpretation, and query execution, using the following capabilities:

Natural-language question: You ask an exploratory question in plain language (for example, "How many sales did we have this month for all cold weather products?"). End users aren’t expected to know the table or project to use.
Semantic discovery: Gemini searches Knowledge Catalog to locate candidate data across your estate, considering key attributes like usage patterns.
Context mapping: The system uses technical metadata, business glossaries, and semantic models to map everyday terms into exact database structures, such as resolving "cold weather products" to specific categories like sweaters, coats, and jackets in an ecommerce catalog.
Permission-aware execution: Gemini generates a SQL query and runs it through the connector under your user credentials, respecting your existing row- and column-level database permissions.

Step-by-step flow of a user prompt from Gemini to query execution.

To accurately navigate an enterprise data estate, language models need business context alongside raw database access. Knowledge Catalog provides a curated context bundle, including technical metadata, business logic, table relationships, and verified queries, that improves query accuracy while reducing latency and token consumption.

Navigating the context graph visually in Knowledge Catalog.

Enterprise analytics has long been fragmented by physical borders: data locked in proprietary silos, open lakehouse tables in cloud object storage, and datasets distributed across multicloud environments. Beyond native BigQuery storage, the BigQuery connector brings conversational analytics to a borderless lakehouse, unifying these environments into a single conversational interface to query open table formats (including Apache Iceberg, Delta Lake, and Parquet) and cross-cloud datasets in AWS or Azure using BigQuery. The connector inherits fine-grained access policies from BigQuery and Knowledge Catalog, so row- and column-level access controls apply whether you query native tables or open lakehouse files.

Gemini accesses data across BigQuery and AWS S3 (through the AWS Glue federated catalog) to answer questions in plain language.

Choosing between connectors and data agents

In addition to using direct Data Cloud connectors to query data in plain language, you can also use Conversational Analytics agents to enable these experiences. While Data Cloud connectors allow you to provide exploratory and ad hoc analysis for users, conversational analytics agents are scoped agents that are built and certified directly by data teams within your organization. These can be built across Data Cloud products such as BigQuery, Looker, and operational databases including Spanner, AlloyDB, and Cloud SQL.

These pre-configured agents allow users to run highly governed, multi-step strategic analyses that are scoped to particular subsets of data, supporting compliance requirements and predictable results for reporting. These data agents can also be published in Gemini for consumption or made available via API for building your own conversational apps in the enterprise.

These approaches are complementary. Gemini direct connectivity also surfaces relevant published agents created by central data teams, so users can decide whether to proceed with ad hoc analysis using the connectors or use the scoped agent. You can choose the right approach for each workload.

Built-in access controls and governance

Data Cloud connector queries execute under the end user’s identity. An employee without access to executive salary data or protected healthcare records cannot query that data through natural-language prompts. The agent only discovers metadata in Knowledge Catalog to which you have access, and only executes queries that your account has permission to run.

How to create Data Cloud connectors

Administrators can configure Data Cloud connectors in the Gemini admin console. From there, admins set up default read-only connections to datasets in BigQuery, operational databases, or Looker, with Knowledge Catalog enabled automatically. Once configured, end users can enable the connector in Gemini to start querying data and building reports in plain language.

Create a connected data store in Gemini and authorize it on first use.

Setup documentation is available for the BigQuery, Spanner, Cloud SQL, and AlloyDB for PostgreSQL connectors, with Looker also in preview.

Unlike data ingestion mode, which copies data into a Gemini data store, Data Cloud connectors use federated query mode to query your data in place without copying or indexing it — using Knowledge Catalog automatically to search accessible metadata and look up business context.

Why use Data Cloud connectors over custom MCP

Compared to connecting to public or self-managed Model Context Protocol (MCP) servers, first-party Data Cloud connectors provide two main advantages:

Built-in skills for higher response quality: Data Cloud connectors include pre-configured agent skills tailored to each data engine, such as advanced SQL and AI function generation for BigQuery and operational databases, to improve query accuracy and response quality.
Guided setup and managed authentication: First-party connectors provide a guided setup experience with managed authentication for Google Cloud databases and BigQuery, removing the need to deploy MCP middleware or manage your own OAuth client configuration for Google Cloud services. In addition, connector-specific settings, such as specifying a default Google Cloud billing project for BigQuery query execution, help administrators enforce governance and manage query costs.

Get started today

Data Cloud connectors help teams find answers from enterprise data sources without waiting on backlog queues. To get started, review the Data Cloud connectors documentation or explore Conversational Analytics agents.

Gemini Omni 1.2 Flash: Fast, cost-effective, and reliable video generation and editing now generally available

Thu, 05 Nov 2026 17:00:00 +0000

Today, Gemini Omni 1.2 Flash is generally available on Google Cloud. Built for the everyday videos businesses make in volume — like product videos, virtual try-ons, and ads — this model is fast, cost-effective, and follows your image, video and audio references more closely than before.

Business users can create with it in Gemini for business and Google Vids, and developers can build with it via API.

Here's what's new:

A lower price: $0.07 per second of 720p video, 30% less than Gemini Omni 1.1 Flash, with similar savings at other resolutions.
Longer clips: Generate and edit videos of up to 15 seconds, up from 10s.
More consistent results: stronger adherence to image and video references keeps products, characters, and brand details consistent from shot to shot.
Faster drafts, in higher resolution: create drafts at 360p or 720p, then upscale the best creations to 1080p or 4K.
More natural speech: You can generate scenes with more fluid, natural-sounding dialogue, clearer lip-sync, and distinct multi-speaker conditioning.
Editing in conversation: refine the same clip, such as swapping a product, changing lighting, or changing camera angle, instead of starting over.

Note: Audio references are gated behind allowlisting.

Gemini Omni 1.2 Flash is ready for production

Gemini Omni 1.2 Flash is production-ready and backed by Google Cloud’s enterprise-grade infrastructure and security, offering Provisioned Throughput. Outputs carry an imperceptible SynthID watermark.

Video creation inside the tools people already use

Design and creative platforms are building Gemini Omni 1.2 Flash directly into their products, so their users can create and edit videos where they already work. Here’s how customers are already using Gemini Omni 1.2 Flash:

“At Artlist, we’re focused on giving creators the control and flexibility to bring their ideas to life. Gemini Omni 1.2 Flash brings exciting capabilities in reference-based generation, video extension, and creative consistency, helping our customers move from experimentation to professional content production at scale.” – Roee Peled, CPTO, Artlist.

^{Artlist uses Gemini 1.2 Omni Flash to bring creative ideas to life}

“Creative teams need models that they can rely on consistently. Gemini Omni 1.2 Flash edits exactly what you ask for, a colorway or a background, and leaves the product untouched, even from detailed images and finished video. Its video-to-video takes footage into entirely new styles.” – Andrew Taco Zhang, Product Marketing Manager, FLORA.

^{FLORA uses Gemini Omni 1.2 Flash to create new styles}

“We ran Gemini Omni 1.2 Flash through our evals at invideo. Honestly, it's really good. It sticks to the prompt, and the physics and sound feel real. If you need fast video at scale, it's one of the best we've tested. And 15 seconds means a full scene in one go.” – Shivam Rastogi, VP of Engineering, invideo.

^{invideo uses Gemini Omni 1.2 Flash to create full scenes in one prompt}

“We’re really impressed by how fast and accurate this new model is. The outputs retain an amazing level of detail, and the model does a great job of making only the changes we ask for while keeping everything else consistent. We’re excited to use it for character and scene replacements, as well as other creative workflows.” – Numan Khan, Creative Specialist, Krea.

^{Krea uses Gemini Omni 1.2 Flash for their creative workflows, from character to scene replacements}

"Gemini Omni 1.2 fits the way our teams work. Draft mode lets us iterate concepts in minutes, then render in 4K for TV-ready AI dramas. Text stays crisp across frames, characters look and sound real, and the speed keeps production moving. Now that it is generally available, we can bring this quality to over 100M Picsart creators." – Narek Hayrapetyan, Head of Product and Partnerships, Picsart.

^{Picsart uses Gemini 1.2 Flash to iterate concepts in minutes}

Runway empowers creatives with advanced AI video tools to produce high-quality content. For organizations, Runway’s enterprise solutions deliver robust collaborative tools to scale professional production. Here is what they created with Gemini Omni 1.2 Flash.

^{Runway uses Gemini Omni 1.2 Flash to produce and scale high-quality content}

Product and listing videos at catalog scale

Retailers and marketplaces use Gemini Omni to turn the photos they already have into short product spins, lifestyle clips, try-on videos, and video ads.

“Photoroom turns product photos into video ads that sell in minutes. We chose Gemini Omni Flash to power our Video Ad Maker and Product Animator because it balances quality, speed and cost. Version 1.2 brings 15-second videos, a top user request. With general availability and SLAs, we can rely on it in production for our sellers and enterprise customers.” - Matt Rouif, CEO and Co-founder, Photoroom.

^{Photoroom uses Gemini Omni 1.2 Flash to power their Video Ad Maker and Product Animator}

“QuickFrame AI helps marketers and business owners create, edit, and launch professional-quality ads for CTV and social in minutes. With Gemini Omni 1.2 Flash now supporting 15-second video generations at 1080p, users can create a complete CTV ad in a single generation, making TV ad creation simpler for businesses of all sizes.” - Richard Girges, Chief Technology Officer, QuickFrame AI.

^{QuickFrame AI uses Gemini Omni 1.2 Flash to create complete CTV ads in a single generation}

Create in Gemini for business, Google Vids and build with the API

Coming soon to Gemini, business users will be able to create a short video from a prompt or a product photo, right in the chat. In Google Vids, teams can combine generated clips with avatars, music and voiceover to create product explainers, teasers, social posts and sales pitches. Businesses can also embed the capabilities of Gemini Omni 1.2 Flash into their applications and agents via the API.

Get started with Gemini Omni

We'll keep improving Gemini Omni based on your feedback, so tell us what you build.

For business users: Try it in Google Vids; coming soon to Gemini for business
For developers: Try it in the Google Cloud console, build with the API, and see the pricing

How Nammu21 automates syndicated loans on Google Cloud Universal Ledger

Tue, 03 Nov 2026 18:00:00 +0000

The syndicated loan and private credit markets, both critical components of corporate finance, have long been constrained by manual processes, leading to multi-week settlements, valuation opacity, and trade friction.

To solve these long-standing challenges, fintech innovator Nammu21 partnered with Google Cloud on a new proof-of-concept. Using its AI-powered platform to structure legal agreements and instantiate them as smart contracts on the Google Cloud Universal Ledger (GCUL), Nammu21 automates the entire loan lifecycle, paving the way for improved settlement times and greater market efficiency.

Digitizing the loan market

While loan markets provide billions in capital to businesses worldwide, their underlying infrastructure has not kept pace with the speed of digital finance. This landscape requires continuous coordination across facility agents, borrowers, and diverse lender syndicates, while also managing highly customized credit agreements with dynamic tranches and complex covenants.

In particular, relying on fragmented, manual interpretation of legal text often results in significant operational drag and data silos. According to the Loan Market Association, secondary market settlement times for par/near par loans in Q4 2024 still averaged 42 days, with a median of 27 days. These notoriously slow timelines tie up capital, increase counterparty risk, and create a heavy reconciliation burden for participants.

Nammu21 transforms bespoke credit documents into programmable assets at scale by rendering each facility's governing terms as a Digital Credit Instrument (DCI). The DCI maps every term — including conditions, consents, eligibility criteria, and operational mechanics that determine what is permitted under a facility — directly back to its originating legal provision. These elements are then expressed as executable behaviours against a loan's live state.

To instantiate these instruments, the proof-of-concept utilizes the secure, high-performance execution environment of GCUL, an institutional-grade, permissioned distributed ledger that helps automate complex financial workflows to deliver near-instant transaction finality and predictable costs without the volatility or security concerns of public blockchains.

A purpose-built stack for digital credit

Combined with Nammu21’s data intelligence, GCUL provides the foundational infrastructure for converting legal document logic into live, programmable contracts. Designed to handle the privacy, scale, and multi-party coordination required by institutional finance, GCUL provides a permissioned, high-performance distributed ledger that delivers deterministic, instant transaction finality for high-value financial settlements.

Programmable contracts on GCUL are written in gculpy, GCUL's domain-optimized programming language. As a strict, statically-typed subset of Python, gculpy is designed to be safe, readable, and auditable, making it ideal for processing complex financial logic.

For regulated financial institutions, gculpy offers distinct operational advantages over other ledger programming models. The language simplifies development by removing the need for complex memory management or workarounds to meet highly restrictive call stack limits, allowing engineers to focus on the financial logic itself.

GCUL also provides robust, built-in permissioning, requiring explicit user authorization before a contract can store data on their behalf. This solves the "unsolicited asset" problem — common on public networks — where unwanted tokens can be sent to an organization's wallet, creating compliance and accounting issues. With GCUL, institutions can maintain full control over which assets they interact with to support their high-value, highly regulated use cases.

Nammu21 serves a dual role in this architecture. First, the platform renders the DCI programmatically from the original legal agreement without manual document reading to establish the terms. Next, it derives gculpy contracts that define the structure of the obligation and govern the loan on GCUL, using the DCI to supply the terms that determine what the loan documents permit, require and block.

Technical architecture: The loan lifecycle on GCUL

Nammu21’s loan implementation on GCUL demonstrates the value of creating a single, immutable source of truth. The proof-of-concept modeled a $1.945 billion Term Loan B — one segment of a broader multi-facility transaction — where the agent bank, borrower, transferor, and transferee all worked off a permissioned view of the same on-ledger rendering.

The lifecycle begins when an agent bank selects a specific credit agreement transformed by Nammu21. The platform instantly provides the structured terms from the document, which are used to instantiate a unique gculpy contract on GCUL, establishing a verifiable, on-ledger digital twin of that specific loan facility.

Next, the agent bank initiates the onboarding of the borrower and lenders. This multilateral action requires that each participant review the request and then provide explicit consent by approving the onboarding invitation from their own interface. Once all parties are onboarded, the agent bank records an immutable hash of the loan's key terms on-ledger, attesting that the on-ledger data matches the canonical off-ledger records.

Subsequent lifecycle events, such as secondary market transfers or repayments, are also managed as multi-step, automated workflows governed by a consent-driven model. For example, a transfer of a loan position requires new lender, borrower, and agent bank consent before it can be executed atomically on the ledger. All payment flows, from interest payments to loan close-outs, settle directly on GCUL using tokenized commercial bank money, eliminating the need for external payment rails.

Transacting with familiar, regulated money on a programmable ledger provides an architecture purpose-built for institutional finance, grounded in three core pillars:

1. Compliance via gated execution: Nammu21’s gculpy contracts instantiated with governing terms, validate every action against the terms transformed from the agreement before execution, mitigating potential covenant breaches or invalid transfers upfront while eliminating the risk of silent failures.

2. Role-based privacy: GCUL enforces strict, role-based access, ensuring that trading strategies, positions, and sensitive loan data remain confidential and are visible only to the necessary counterparties in a transaction.

3. Settlement finality: Every on-ledger action completes near-instantly and anchors on canonical data derived from the original credit agreement, replacing provisional, reconciliation-heavy models with real-time, trusted verification.

The business impact of moving on-ledger

This proof-of-concept illustrates a structural shift in how credit markets can operate. When each facility's governing terms are rendered as a DCI and executed on GCUL, the operational mechanics of lending markets change fundamentally, rather than incrementally.

Values can be reconciled bilaterally, but governance logic cannot – and the market’s current friction stems directly from each party's independent interpretation of the rules. Resolving this issue on-ledger is what enables settlement times to compress, back-office overhead to shrink, and counterparty risk to be mitigated by design rather than by manual checking.

This same architecture can be applied wherever credit terms are bespoke and manually interpreted. By freeing up capital that would otherwise be trapped in lengthy settlement cycles, new liquidity is unlocked and a more dynamic and efficient secondary loan market can emerge.

^{The team would also like to thank Someera Khokhar, Jordan Shelley, and Saskia A. Aalders at Nammu21 and Googlers Ishan Singh, Kathryn Stavish, Tom Chudy, and Stathis Onasoglou for their contributions.}

Introducing Physical-Engineering’s Last Exam for agents

Tue, 03 Nov 2026 08:00:00 +0000

Authors: Burak Gokturk, VP Research & Development; Rujun Han, Research Scientist; George Lee, Product Manager.

Today, the Google AI & Infrastructure team is announcing Physical-Engineering's Last Exam (PLE), a new benchmark and open competition, developed in collaboration with Mercor, and built to test a frontier model’s ability to complete tasks in mechanical, electrical, civil and structural engineering.

To help the research community measure and improve how agents handle physical engineering work, we are releasing 29 public tasks, their grading tools, and an open competition leaderboard at ple-bench.com.

Frontier labs and the broader research community use model examinations to test AI’s progress in completing tasks in a specific discipline. Humanity's Last Exam measures a model’s general knowledge. Agents' Last Exam measures a model’s ability to complete digital professional work. Physical-Engineering's Last Exam (PLE) measures a model’s ability to complete work that helps build the physical world. Physical engineers design skyscrapers, power grids, and computer chips we rely on, yet we could find no existing benchmark that tests whether frontier models can do that work end to end.

Why textbook benchmarks fall short for physical engineering

Frontier language models now score above 90% on many academic benchmarks. However, those benchmarks mostly test textbook-style questions, single tool calls, or small, self-contained design problems. For AI developers and researchers building agents for physical engineering, existing benchmarks do not show whether an agent can produce safe, code-compliant, and cost-effective designs under a licensed engineer's review and control. As the demand for infrastructure development outpaces financing, there is a growing opportunity for agents to support engineers working to close this gap.

Inside PLE: Real-world tasks built with practicing engineers

PLE closes this evaluation gap with 85 long-horizon tasks across four core disciplines: mechanical (30 tasks), electrical (27 tasks), civil (18 tasks), and structural (10 tasks) engineering.

Instead of answering a single prompt, an agent works inside a containerized project folder — built on Mercor’s Archipelago environment — that mirrors what a practicing engineer receives on the job. Each task holds an average of 61 files (5,170 total files across the benchmark), including PDFs, spreadsheets, emails, 2D CAD drawings, 3D BIM models, and document revisions. The agent must reason through these files, find the governing building codes and physical limits, run calculations, and generate production-ready deliverables that a licensed reviewer would review and approve.

In collaboration with experts from Mercor and licensed practicing engineers from industry, we built PLE using two methods:

Expert-authored tasks (44 tasks): Practicing engineers created the project folders, prompts, and grading rubrics based on real professional engineering tasks.
Co-authored tasks (41 tasks): Practicing engineers designed the initial task and an agent pipeline assembled the task's environment and grader from licensable engineering standards. Engineers then validated and increased the difficulty of every task assembled by the agent.

All frontier models score below 50%

We tested 14 frontier models on PLE using seven publicly available agent harnesses (Antigravity, Claude Code, Codex, Hermes, OpenCode, OpenHands, and Pi). Because PLE evaluates both code compliance and cost efficiency, agents only earn points if their design passes every hard safety check first. Our baseline results show a wide gap between current models and real-world engineering requirements:

High headroom: All frontier models score below 50% on our private test set, with the highest-performing model on its native harness reaching a mean score of 48.3%.
Hard safety gates: While the top-scoring agent configuration satisfied 91% of individual constraints on co-authored tasks, only 41.3% of its deliverables passed every hard safety and feasibility check required.
Model vs. harness impact: Choice of model drives performance far more than the harness (up to a 50-point spread across models, versus a median 3-point spread across harnesses for the same model), though harnesses significantly affect run cost.

Test your agent and join the open competition

Evaluating AI agents on realistic project files, strict safety rules, and material costs is a practical step toward building helpful tools for physical engineers. Here is how you can get started with PLE today:

Check out the full technical breakdown in our PLE paper on arXiv.
Download the 29 public tasks, grading scripts, and evaluation container from our GitHub repository.
Bring your own agent setup and submit it for evaluation on our private test set at ple-bench.com—with 13 Google-sponsored models running free on our servers and prize details coming soon.

What Google Cloud announced in AI this month

Mon, 02 Nov 2026 16:00:00 +0000

Editor’s note: Want to keep up with the latest from Google Cloud? Check back here for a monthly recap of our latest updates, announcements, resources, events, learning opportunities, and more.

This October, we took massive leaps forward in delivering flexible, open, and powerful enterprise AI solutions. From bringing Gemini 4 Argon to Google Cloud to expanding choice and flexibility, we’re giving organizations the tools they need to bring complex, multi-agent AI workflows into production safely.

Here’s a quick look at the major milestones, expert insights, and practical developer guides to read through and save.

Top announcements

Gemini 4 Argon available on Google Cloud, built to help solve your business’ hardest problems: We trained Argon to work with both depth and precision, which allows it to maintain context across long, multi-step tasks which are reinforced with strong domain expertise along the way. Whether your team is running complex software builds, handling detailed financial and legal work, or deploying cybersecurity tools like CodeMender, Argon helps your teams maintain focus from start to finish.
Expanding Claude availability and deployment choice on Google Cloud: We announced the general availability of an Anthropic-managed deployment option for Claude Models on Google Cloud. This gives teams more choice in how they run their models, helping them consolidate AI workloads on Google Cloud alongside our existingGoogle-managed deployment option.
Google Cloud expands model choice and flexibility, bringing OpenAI models to customers: Google Cloud is now the only hyperscale cloud provider offering access to all three leading closed source LLMs, alongside more than 200 first-party, third-party, and open models on a unified enterprise AI platform.
Claude Opus 5.5 and Claude Sonnet 5.5 are now available in Google Antigravity: In keeping with Google Cloud's commitment to an open ecosystem and to give developers the right model for every job, Claude Opus 5.5 and Claude Sonnet 5.5 are now available in Google Antigravity for Google Cloud customers. Claude Opus 5.5 is built for complex work while Claude Sonnet 5.5 is a faster, lower-cost complement.
Introducing Gemini for government services: Today we're bringing the Gemini agent to public service with Gemini for government services in preview, part of our new suite of purpose-built industry solutions. It adds skills built around government work, connections to the systems agencies already run, and specialized partner agents, with governance running underneath it all.
Gemini Robotics Embodied Reasoning 2 is now available on Google Cloud: For the first time, organizations can pair Google DeepMind’s latest embodied reasoning model for robotics with the enterprise-grade security, multi-agent orchestration, and end-to-end lifecycle management of Google Cloud — making it easier to simulate, evaluate, and deploy safe, adaptive physical AI at scale.

Thought leadership (editor’s pick)

A new gear: How startups are advancing robotics with Google AI: We marked a major milestone in physical AI with the launch of Gemini Robotics Embodied Reasoning 2 (ER 2) on the Gemini platform and AI Studio, bringing Google DeepMind’s frontier multimodal reasoning and long-horizon task planning for robotics to Google Cloud. Read more from Darren Mowry, VP of Global Startups and Investor Ecosystem at Google, about the exciting work these startups are undertaking.

News you can use

Developer’s guide to implementing MCP apps in Gemini: Enterprise assistants still respond mostly in text. However, that can break down for visual, multi-step work, where every adjustment costs another slow chat turn. This guide takes you through implementation in Gemini.
Four evaluation engineering lessons learned from building agent plugins: Here are four eval lessons we learned from building the new plugin that you can apply to your own builds.
Who judges the judge? Three helpful ways to audit your LLM evaluators: In this post, we explore what’s holding enterprises back from strong evals, and share an easy three-pillar method you can use to improve your quality.

Stay tuned for monthly updates on Google Cloud’s AI announcements, news, and best practices. For a deeper dive into the latest from Google Cloud customers, read our monthly recap, Cool stuff customers built.

aside_block: <ListValue: [StructValue([('title', '$300 in free credit to try Google Cloud AI and ML'), ('body', <wagtail.rich_text.RichText object at 0x7fc4078c6190>), ('btn_text', 'Start building for free'), ('href', 'http://console.cloud.google.com/freetrial?redirectPath=/vertex-ai/'), ('image', None)])]>

September

This September, we brought you new how-to guides, model updates, and security innovations to help you find-and-fix. From Gemini 3.8 Live with Live Avatar to Mantis, there’s no shortage of new ways to apply AI to your work.

Notably, we announced Gemini 4 Argon, our next era of frontier intelligence. Gemini 4 Argon delivers frontier performance in complex workflows across real-world software engineering, enterprise knowledge work like legal and finance, and cybersecurity defense. You can learn more here.

Top announcements

Gemini 3.8 Live with Live Avatar is now generally available: Following our announcement of Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, Gemini 3.8 Live with Live Avatar is now generally available in Gemini Enterprise.

Getting started with Mantis, our open-source bug finding-and-fixing harness: Available to all as an open-source framework, Mantis is part of Google’s internal approach to find and fix vulnerabilities at machine-speed. It creates a more effective scalable, context-aware repository analysis.

Thought leadership (editor’s pick)

Racers, start your agents: Formula E brings realtime AI to the edge: Chasing records at the Goodwood Hillclimb, Formula E used Gemini Enterprise, Gemma, and a Pixel phone bolted to the dash to push the envelope of AI and edge computing.

News you can use

Best practices guide for customizing Gemini models via Reinforcement Learning (RL): In this guide, we will walk through practical best practices for using RL fine-tuning service. We'll start with a short tour of the RL training loop, how to decide if and when to use RL, and introduce how to get the most value from this approach.

Empower your agents with the Google Cloud CLI remote MCP server: Powered by the popular gcloud and bq (BigQuery) command-line tools, this new server gives AI agents immediate, broad access to command-line operations for managing Google Cloud infrastructure and working with advanced BigQuery workflows securely and seamlessly.

Graph Workflows in ADK: Everything you need to know: Using a refund workflow in ADK, we'll cover fan-out and fan-in, deterministic and agent routers, human-in-the-loop pauses, parallel workers, and dynamic orchestration— along with when to use a static graph or let Python decide what runs next.

August

This month, we focused on making AI highly practical for your business, including its associated costs. This meant tailoring our models for specialized industries – starting with Financial Services and Legal – and helping you keep your budgets under control. Let’s dive in!

Top announcements

FinOps for the AI era: New flexible billing and cost controls for agents: To help you get better return on AI, we announced expanded billing flexibility and new cost management tools for agent workloads across Gemini Enterprise and developer tools like Google Antigravity in Gemini Enterprise and Android Studio.
Gemini Enterprise for Financial Services: We’re bringing Google’s agentic AI directly into the workflows of capital markets and corporate banking.
Gemini Enterprise for Legal: Gemini Enterprise for Legal provides an integrated, fully governed environment configured for rapid deployment across firms and corporate legal departments.
Expanding Google Antigravity for enterprise customers: Antigravity is available now as part of eligible Gemini Enterprise app subscriptions, including out-of-the-box administrative and spend controls.

Thought leadership (editor’s pick):

Tokenomics: Why smart teams spend more on AI, on purpose: Hear from Eric Lam, Head of Value, Delta, Google Cloud Consulting about how disciplined organizations are moving past reactive sticker shock over AI bills and embracing tokenomics.
Meet the researcher fighting AI hallucinations at Google Cloud: Cyrus is a senior research scientist at Google. Lately, his focus has shifted to large language models, specifically a persistent issue known as hallucination, which is when an artificial intelligence model lacks the correct facts but confidently invents an answer anyway.
What sports cars can teach us about optimizing AI spend: More tokens doesn't always mean better AI. Read our conversation with Mike Clark, Director of Product Management for Gemini Enterprise Agent Platform, on how to balance horsepower with efficiency and get the highest return out of every dollar you spend on AI.

News you can use:

Looking for a steer on your foundation or inspiration for your next project? Take a look at some of our favorite how-to guides from August:

10 questions every startup should answer before moving to production with their AI prototype: These ten are scoped to the prototype-to-production transition itself. Each question ends with a short, runnable snippet you can copy into your own project today. Adjacent decisions that matter just as much but aren't specific to that move, your data layer and RAG architecture, CI/CD, network design, are deliberately out of frame here.
Your chance to start building AI agents from the absolute basics: Agent Valley is a free, 5-week live learning series designed to take you from scratch to building your very own hands-on agent systems. And instead of staring at boring terminal lines, you’ll be building and playing inside a tiny, low-poly virtual world!

July

Since launching the Gemini Enterprise Agent Platform a few months ago, we’ve watched businesses move from basic experiments to serious, production-grade builds. We want to make it even easier — and more secure — for you to scale those systems.

Along with a batch of new platform updates, this month we’ve put together 13 practical demos and 20 diagnostic questions to help your engineering teams align on a strong architectural blueprint. Let’s dive in!

Top announcements

What’s new in Gemini Enterprise Agent Platform: In this helpful recap, we announced some of our most popular capabilities are available for everyone, from Agent Runtime to Agent Identity.
Now in preview: Find and fix software vulnerabilities with CodeMender: As adversarial AI threats accelerate attacks on code, security teams must counter them with machine-speed defenses that can automate code remediation and fight AI with AI. You can learn more about CodeMender and review the documentation here.
Solve harder problems with AlphaEvolve, now available to everyone on Google Cloud: AlphaEvolve is a code optimization and discovery agent built on top of Gemini that helps solve the hardest algorithmic problems and achieve breakthroughs for your business and research.

Thought leadership (editor’s pick):

If automation requires delegation, then delegation requires trust. But letting an AI agent run on its own is a big leap for any business. While the productivity gains are clear, the fear of losing control is very real. This month, we sat down with our experts to discuss how leaders can navigate this shift by focusing on transparency, predictability, and setting clear boundaries for how agents handle weird data exceptions.

Here’s our picks for the month to learn more.

How leaders can scale AI by trading control for trust (Q&A): We sat down with Michael Gerstenhaber, VP of Product Management for Gemini Enterprise, to discuss why the future of AI is about defining safe boundaries.
What makes an AI agent trustworthy: Context is fast becoming one of the most valuable assets a company owns. Prajakta Damle, Senior Director, Product Management, shares what it takes to get trustworthy AI right.

News you can use:

What if you’re looking for a steer on your basic foundation, inspiration for recipes, or some inspiration? Take a look at some of our favorite how-to guides from July:

Automate your agent development lifecycle using any coding agent: Stuck prototyping? With Agents CLI skills, you can go through the different phases of the entire agent lifecycle without ever leaving your coding agent.
Why AI apps fail in production (And how Google solved it): Only 5% of AI prototypes make it to production, and the other 95% fall into the validation abyss. How can you move confidently into production?
13 hands-on demos to build on Gemini Enterprise Agent Platform: Not sure where to start with Agent Platform? Here’s 13 ways you can stir up some creativity.

June

Our main focus in June was helping your teams build, scale, and secure AI. Today, we’re sharing a fresh roundup of updates designed to help you run smarter, more secure applications while keeping everything under your control.

We even shared a cool virtual shopping demo at Cannes to show how retailers can make product discovery more exciting. Let’s dive in!

Top announcements

Introducing the Open Knowledge Format: We introduced the Open Knowledge Format (OKF), an open specification that formalizes the LLM-wiki pattern into a portable, interoperable format. This is a vendor-neutral, agent- and human-friendly standard for representing the metadata, context, and curated knowledge that modern AI systems need.
Collaboration with Apple on its expanded Private Cloud Compute (PCC) systems: Our collaboration with Apple is built on a foundation of deep commitment to privacy that leverages Google Cloud's security and privacy technologies. At the heart of this collaboration is our Confidential Computing portfolio and our Titanium security architecture.
Claude Fable 5: Available on Google Cloud: Claude Fable 5, Anthropic’s latest frontier model, is now generally available on Google Cloud. This launch is the latest proof point of our ongoing commitment to bring the industry's latest models straight to our Agent Platform.
Cloud Atelier: How Gemini Enterprise is helping restyle the retail playbook: This year at Cannes, we showcased Cloud Atelier — a destination-based, virtual shopping experience that highlights how retail brands can turn this classic dilemma into an exciting moment of product discovery.

Thought leadership (editor’s pick):

How Google Cloud Security uses AI internally: To counter machine-speed, AI-driven threats, we’ve worked hard to transition Google Cloud’s security posture to an autonomous, proactive model. By embedding specialized AI agents directly into our software development lifecycle (SDLC), we’ve created automated guardrails that protect code at a scale and speed unreachable by human teams — and we’re taking steps to make those same guardrails widely available.
The 4 lessons that guided AI Threat Defense: We introduced Chris Betz as the new CISO of Google Cloud. For his first Cloud CISO Perspectives, Chris shares four key lessons we learned about using AI to the defender’s advantage while building AI Threat Defense.

News you can use:

5 lessons from red teaming AI applications: To help you build AI securely, Mandiant has developed a proactive, risk-based approach centered on the Good AI Assessment (GAIA) Top 10, outlined in our new report, Secure Development of Generative AI Applications: A Proactive Approach.
How to unlock true ROI in software development – a deep dive into the latest DORA research: To help you evaluate the costs and business benefits of AI, we recently shared the DORA: ROI of AI-assisted software development report. This research offers a practical approach to help your team work through early adoption challenges, align engineering plans, and drive business growth.
Agent Factory Recap: 100X engineering with AI agents in Google Antigravity 2.0: In this episode of the Agent Factory, Shir Meir Lador, Head of AI Engineering, Google Cloud Developer Relations, sat down with Rody Davis, one of Google’s top agentic engineers. They dive into the massive shift from traditional IDEs to agent-first platforms, the reality of code reviews in an AI-driven world, and how to use "skills" to perform at a 100X level.

May

We’ve had a busy month! Between announcing Gemini Spark and Gemini 3.5 at Google I/O – and unveiling Google AI Threat Defense, our latest AI-powered cybersecurity solution, we had a lot to share with Google Cloud customers. Keeping up with the latest news takes time, so we gathered the most important announcements, thought leadership, and technical guides in one place to help you quickly catch up.

To learn more about our I/O announcements, here’s everything you need to know for Google Cloud customers, and top news for startups.

Top announcements

Introducing Google AI Threat Defense to help you outpace the adversary: Google Cloud is introducing a comprehensive AI-powered cybersecurity solution — Google AI Threat Defense — an always-on autonomous security platform. Learn more here.

Gemini 3.5: Our latest family of models combines frontier intelligence with action – starting with Gemini 3.5 Flash.
Gemini Omni: Our new model is a leap forward in world understanding, multimodality, and editing, letting you generate any output from any input, starting with video.
Google Antigravity: Google Antigravity’s expanded capabilities and new integration with Agent Platform bring agentic development to your entire organization.
Gemini Spark: For Gemini Enterprise and Workspace customers, Gemini Spark is your 24/7 personal AI agent that helps you work more efficiently by autonomously taking action on your behalf, under your direction.
Google Workspace: Google Pics, our new image generation and editing tool, and new voice features in Gmail, Docs and Keep, help reimagine how you work.
Managed Agents API on Agent Platform: Allows developers to build and run custom agents inside secure, Google-hosted environments that seamlessly integrate with Agent Platform.
CodeMender: A powerful AI security agent provided through Agent Platform, CodeMender can help find and fix vulnerabilities in your code.

Nano Banana 2 and Nano Banana Pro are generally available: Available today via Gemini Enterprise Agent Platform, organizations are already putting the models to work. Learn more here.

Thought leadership (editor’s pick):

Cloud CISO Perspectives: How Google + Wiz changes multicloud strategy for CISOs: Vinod D’Souza, director, Office of the CISO, shares highlights from his RSA Conference fireside chat with Anthony Belfiore, chief strategy officer, Wiz. While threat actors have seen gains from the adversarial misuse of AI, Google and Wiz are tackling these challenges head-on by combining Wiz's deep cloud telemetry with Google's world-class AI and quantum research to help CISOs and their organizations meet the needs of the agentic enterprise era. Read more here.

News you can use:

What Google I/O '26 means for developing agents on Google Cloud: Dig deep into how Gemini Enterprise Agent Platform and the new developer tools shared at I/O fit together, unpack the spectrum of choice for building, and share what we’d actually try first. Learn more here.

Five must-have guides to move agents into production with Gemini Enterprise Agent Platform: Here is a look back at our five-part series covering the architecture patterns and best practices you need to move your agents into production. Learn more here.
How to build an AI-ready security program for the public sector: From industrial control systems to decades-old municipal databases, here’s our CISO guidance to prep AI-ready security programs for the public sector. Learn more here.

April

We hosted Google Cloud Next in Las Vegas on April 22, announcing incredible innovations from Gemini Enterprise Agent Platform to our eight-generation TPUs. We also expanded the Gemini Enterprise app in collaborative ways – now, with new features like Projects, you can work side-by-side with your agents and colleagues.

If you missed the livestream, take a look at our Day 1 recap. It’s been incredible to see how customers have been applying AI in thousands of ways — so far, we’ve counted more than 1,300 examples.

Top announcements

1. Gemini Enterprise Agent Platform: Our new, comprehensive platform to build, scale, govern, and optimize agents. Moving forward, all Vertex AI services and roadmap evolutions will be delivered exclusively through the Agent Platform, rather than as a standalone service, to power the next generation of agent development.

The platform is designed around four core pillars — build, scale, govern, and optimize — that allow teams to collaborate seamlessly. Learn more about Agent Platform here.

2. Gemini Enterprise app has all the key components to let teams discover, create, share, and run AI agents in a single environment. At Next ‘26, we introduced several new capabilities in the Gemini Enterprise app:

Agent Designer uses the same no-code agent designer experience of Agent Platform and lets employees build sophisticated schedule- and trigger-based agents using any enterprise connector. It gives you a virtual flowchart of your agent, allowing you to inspect, test, and approve workflows, ensuring total transparency for executing critical business processes.
Long-running agents are designed to execute complex business processes. They can work autonomously in secure cloud sandboxes, giving agents the ability to orchestrate business logic, write code to build custom tools, and complete multi-step work like reconciliation activities or sales prospect sequencing — without needing constant prompting.
Inbox in Gemini Enterprise provides a central location to monitor, guide, and help manage all of your agent activity, including your long-running agents. Notifications are intuitively categorized into actionable groups like "Needs your input," "Errors," and "Completed.”
Projects create a dedicated space where the agent’s memory is confined to the files and conversations your team adds. By connecting it to data sources including Google Drive, NotebookLM, and Google Group Chats, the agent becomes an expert on a specific topic and can provide team members daily briefings or status updates without digging through months of documents.
Skills create simple shortcuts using an “@” mention for repetitive tasks such as applying brand guidelines, formatting a report, and accessing specific data.
Canvas gives our customers an interactive editor directly within Gemini Enterprise. It allows teams to easily create and edit Docs and Slides, and even export to Microsoft 365 files, within the same experience.
Agent Gallery provides access to third-party agents from partners like Adobe, Atlassian, Lovable, and ServiceNow, and is adding more third-party connectors for Asana, Mailchimp, Workday, and more. These integrations enable your agents to retrieve data and execute tasks with your systems-of-record.

3. AI Hypercomputer: Designed specifically for demanding AI workloads, our AI Hypercomputer is an advanced, purpose-built architecture that unites performance-optimized hardware for compute, storage, networking, open software and machine learning frameworks — as well as flexible consumption models — into a single, integrated system. We are announcing innovations at every layer of the AI Hypercomputer:

TPU 8t, optimized for training, uses breakthrough Inter-Chip Interconnect (ICI) technology to scale up to 9,600 TPUs and 2 PB of shared, high-bandwidth memory in a single superpod. It achieves 3x the processing power of Ironwood and delivers up to 2x more performance/Watt.
TPU 8i, optimized for inference, uses our new Boardfly topology to directly connect 1,152 TPUs in a single pod. It features 3x more on-chip SRAM compared to previous versions to host larger KV caches entirely on-silicon and integrates a specialized Collectives Acceleration Engine. Taken together, TPU 8i delivers 80% better performance per dollar for inference than the prior generation, enabling millions of concurrent agents to run cost-effectively.

4. The Agentic Data Cloud: A new data architecture built for the speed and scale of agentic AI. The Agentic Data Cloud delivers an AI-native architecture, allowing agents to perceive, reason, and act on your behalf in real-time, including:

Cross-Cloud Lakehouse, standardized on Apache Iceberg, is our Lakehouse that enables you to leave your data in AWS or Azure (coming later this year) while querying it instantly — without the friction of vendor lock-in or the cost of data movement
Knowledge Catalog constructs a unified, dynamic context graph of your entire business enabling you to ground agents in all of your business data and semantics. With Smart Storage and the Object Context API, files in Google Cloud Storage are instantly tagged and enriched with metadata before an agent touches them. Then our Knowledge Engine uses Gemini to autonomously tag, define logic and instantly map complex relationships across your entire enterprise, providing the semantic definition your agents have been missing.

5. Protecting the agentic enterprise: Security built for the AI era. Our full-stack AI approach, from the chips to the models, gives you a competitive advantage with better integration and velocity to help protect customers. Not only can Google action insights from the world’s largest threat observatory and Mandiant frontline experts, but we also bring cutting-edge insights and breakthroughs from Google DeepMind, to help make your platforms more secure.

Agentic defense: Three new agents in Google Security Operations can help hunt threats, engineer detections, and provide context on third parties. You can build your own security agents with remote Google Cloud model context protocol (MCP) server support for Google Security Operations, now generally available. You can also access the MCP server client directly from the Google Security Operations chat interface, available in preview.
Protecting AI and cloud apps across any infrastructure with Wiz: Newly expanded AI coverage helps build secure agents across clouds and AI studios. New AI-Bill of Materials in development tools can help secure AI-generated code and mitigate the risk of shadow AI. Learn more.
Securing agents and the agentic web: Model Armor can integrate with Agent Gateway, and new Agent Identities provide more layers of defense against shadow AI. Google Cloud Fraud Defense, the next evolution of reCAPTCHA, offers agent-specific capabilities that can help secure the agentic web as well as the entire user and customer journey.
Trusted Cloud: We’re simplifying permissions with modern IAM, and advancing Google Cloud security with new capabilities in Security Command Center plus new innovations in data and network security.
New partner-supported workflows for Google Security Operations: This new robust cohort of partner integrations includes partners developing their own agentic security operations centers (SOCs).

You can catch up on all our security announcements from Next ‘26 here.

News you can use

Guide to prompting Gemini 3.1 Flash TTS (text-to-speech): The new TTS model introduces a high level of controllability by allowing you to steer the delivery using more than 200 audio tags. We'll share how to get strong results from the model, whether you are building accessible gaming soundtracks, banking systems, or audiobooks. Learn more about the model here.
Ultimate prompting guide for Lyria 3 models: Lyria 3, Google's family of music-generation models, is designed to give you granular control over vocals, instrumentation, and arrangement. So we spent weeks testing against every musical genre and use case we could imagine. We put together this guide to share exactly what we learned and how you can get the best results.
How to find the sweet spot between cost and performance: This guide will walk you through Google Cloud's flexible gen AI infrastructure options, showing you how to find that sweet spot on the efficient frontier between cost and performance. We'll start with the foundational pay-as-you-go (PayGo) models and then explore how to layer on more specialized options to build a robust and cost-effective gen AI strategy.
Essential AI and cloud security now on by default: To support the next generation of AI innovators, we are offering on by default essential AI security and cloud security in Security Command Center Standard.
Securing AI inference on GKE with Model Armor: Here’s how to secure AI inference on Google Kubernetes Engine with Model Armor and high-performance storage.
Cloud CISO Perspectives: AI, security, and the workforce of the future: You can’t bring traditional security to an AI fight, so how do we defend against AI-powered attacks, boost defenders with AI, and secure AI use? Drop in on this RSA Conference fireside chat between Francis deSouza, Google Cloud COO and President, Security Products, and Nick Godfrey, senior director, Office of the CISO.

March

March was a busy month for our AI teams. We launched Gemini Embedding 2, rolled out a highly cost-effective Veo 3.1 Lite model, and officially welcomed the Wiz team to Google Cloud to help redefine security in the AI era.

Alongside these launches, we created comprehensive guides to help you get the most out of these models, from prompting formulas for Nano Banana 2, to practical advice for optimizing your TPU training. Here’s a quick look at the latest news and resources to help your team build what’s next.

Top hits:

Gemini Embedding 2: Our first natively multimodal embedding model: Gemini Embedding 2 is our first natively multimodal embedding model that maps text, images, video, audio and documents into a single embedding space, enabling multimodal retrieval and classification across different types of media — and it’s available now in public preview.
Build with Veo 3.1 Lite, our most cost-effective video generation model: This model empowers developers to build high-volume video applications, at less than 50% of the cost of Veo 3.1 Fast, but with the same speed. This rounds out the Veo 3.1 model family, giving developers flexibility based on needs. For Cloud customers, it’s now available on Vertex AI.

Here’s a fun bonus: Check out our ultimate prompting guide for Veo 3.1 to get started.

Welcoming Wiz to Google Cloud: Redefining security for the AI era: Google has completed its acquisition of Wiz, a leading cloud and AI security platform. The Wiz team will join Google Cloud, and we will retain the Wiz brand. With the addition of Wiz, we will provide customers with a comprehensive platform to secure their cloud and hybrid environments, as well as accelerate threat prevention, detection, and response.
Gemini 3.1 Flash Live: Making audio AI more natural and reliable: We’ve improved 3.1 Flash Live’s overall quality, making it more reliable for developers and enterprises to build voice-first agents that can complete complex tasks at scale. On ComplexFuncBench Audio, a benchmark that captures multi-step function calling with various constraints, it leads with a score of 90.8% compared to our previous model.

News you can use:

The ultimate Nano Banana prompting guide: This is a must-read for anyone working with Nano Banana. We spent weeks testing Nano Banana 2 and Nano Banana Pro against every use case we could imagine to test its limits. We put together this guide to share exactly what we learned and how you can get the best results. Here’s an example formula: [Reference images] + [Relationship instruction] + [New scenario]

A developer’s guide to training with Ironwood TPUs: In this guide, we hear from Lillian Yu, CPA, CA , Product Strategy and Operation, and Liat Berry, Product Manager, on five strategies within the JAX and MaxText ecosystems designed to help developers refine training efficiency and hit peak performance on Ironwood hardware.
How to build production-ready AI agents with Google-managed MCP servers: In this guide, we anchor on a specific example. Cityscape is a demo agent built with Google's Application Development Kit (ADK) that turns a simple text prompt — like "Generate a cityscape for Kyoto" — into a unique, AI-generated city image. Check out the guide to learn more.

February

In February, we’re giving developers more reasoning power with Gemini 3.1 Pro and Claude 4.6, and faster creative scaling with Nano Banana 2. We’re also opening up new training programs and step-by-step guides to help you tackle the hardest parts of the AI lifecycle, from capacity planning to mounting defenses against AI-powered attacks.

Here’s a rundown of our latest news, tools, and resources to help you build what’s next.

Top hits

Pro-level image generation gets faster and more accessible with Nano Banana 2: To build creative that stands out, you need models that naturally integrate into your workflows and scale with ease. Check out our blog to see how this comes to life (and how customers are putting the model to work).

Introducing Gemini 3.1 Pro on Google Cloud: Gemini 3.1 Pro is a clear step forward in reasoning, designed to solve tougher problems, giving you the reasoning depth your business needs. Gemini 3.1 Pro is available starting today in preview in Vertex AI and Gemini Enterprise. Developers can access the model in preview via the Gemini API in Google AI Studio, Android Studio, Google Antigravity, and Gemini CLI.
Announcing Claude Opus 4.6 and Claude Sonnet 4.6 on Vertex AI: Now generally available on Vertex AI, explore our sample notebook to get started and visit our documentation for comprehensive pricing and regional availability details.
New AI threats report: Distillation, experimentation, and integration: John Hultquist, chief analyst, Google Threat Intelligence Group, details what security leaders should know from our newest AI threat report on experimentation, integration, and distillation attacks.

News you can use

A developer's guide to production-ready AI agents: To help developers work through these challenges, we've published a collection of guides covering the full agent lifecycle. These resources first appeared during Kaggle’s 5 days of AI Agents Intensive, and they’ve proven so popular and useful, we wanted to make sure a wider audience had access, as well.
Gemini Enterprise Agent Ready (GEAR) program now available: We opened the Gemini Enterprise Agent Ready (GEAR) learning program to everyone. As a new specialized pathway within the Google Developer Program, GEAR empowers developers and pros to build and deploy enterprise-grade agents with Google AI.
Your guide to Provisioned Throughput (PT) on Vertex AI: Check out this deep-dive blog designed to show you the resources available to you today on Vertex AI, and how you can get started capacity planning.
How AI can boost defenders, from defense in depth to the cyber kill chain (Q&A): We know that defenders are also developing powerful AI tools, but what’s still unknown is what it could mean for enterprise software ownership if companies have to constantly mount AI-directed defenses at AI-powered attacks?

Janurary

We used to have to learn the language of computers. In 2026, they’re learning ours.

We kicked off the year by exploring the future of agentic commerce, where AI agents navigate the web to find and buy products for us. Our leaders call this the "invisible shelf" — a world where commerce isn't tied to a specific website. To make this reality scalable, we announced the Universal Commerce Protocol (UCP), a shared language that allows agents and retailers to understand each other.

We brought that same fluency to our creative and technical tools:

Updates to Veo 3.1 allow creators to use simple inputs — like reference images — to generate precise, mobile-ready video.
Natural language queries: With Comments to SQL in BigQuery, we’re removing the language barrier to data. Engineers can now write queries by describing their intent in natural language, prioritizing the question over the code.

Let’s dive in.

Top hits

1. Gemini Enterprise for Customer Experience (CX): Specifically built for agentic retail, this platform transforms fragmented search, commerce and service touch points into one seamless journey — whether you need a shopping assistant, a support bot, agentic search or help with merchandising.

2. We announced Universal Commerce Protocol (UCP): A new open standard for agentic commerce that works across the entire shopping journey — from discovery and buying to post-purchase support. UCP establishes a common language for agents and systems to operate together across consumer surfaces, businesses and payment providers. So instead of requiring unique connections for every individual agent, UCP enables all agents to interact easily. UCP is built to work across verticals and is compatible with existing industry protocols like Agent2Agent (A2A), Agent Payments Protocol (AP2) and Model Context Protocol (MCP).

3. We updated Veo 3.1, including improvements to Ingredients to Video and Portrait mode: Veo is getting more expressive, with improvements that help you create more fun, creative, high-quality videos based on ingredient images, built directly for the mobile format. This includes:

Improvements to Veo 3.1 Ingredients to Video, our capability that lets you create videos based on reference images.
Native vertical outputs for Ingredients to Video (portrait mode) to power mobile-first, short-form video creation.
State-of-the-art upscaling to 1080p and 4K resolution 1 for high-fidelity production workflows.

These updates are launching in the Gemini app, YouTube, Flow, Google Vids, the Gemini API and Vertex AI.

4. Vibe querying with comments-to-SQL: Crafting complex SQL queries can be challenging. Often, engineers simply want to express their data needs in plain English directly within their SQL workflow. That’s why we’re introducing Comments to SQL in BigQuery. This feature makes writing queries using natural language – ‘vibe querying’ – a reality. Learn more in the blog.

News you can use

Mastering Gemini CLI: Your complete guide from installation to advanced use-cases: We’ve teamed up with DeepLearning.ai and are excited to announce a free course – Gemini CLI: Code & Create with an Open-Source Agent. This course isn’t just for developers; we dive into practical use cases for various tasks such as data analysis, content creation, and personalized learning.
How Google SREs use Gemini CLI to solve real-world outages: In this article, we’ll delve into real scenarios that Google SREs are solving today using Gemini 3 (our latest foundation model) and Gemini CLI—the go-to tool for bringing agentic capabilities to the terminal.
Getting started with Gemini 3: Deploy your first Gemini 3 app to Google Cloud Run: In this blog, we will show you how to vibe code your first app—which leverages the Gemini 3 Flash Preview model and deploy it as a publicly accessible URL on Google Cloud Run. Google AI Studio lets you go from idea to app quickly by using natural language to generate fully functional apps using the power of Gemini 3.
Practical guidance: Building with the Secure AI Framework (SAIF) on Google Cloud: We know that security and data privacy are the top concern for executives when evaluating AI providers, and security is the top use case for AI agents in a majority of industries. To help you build AI boldly and responsibly, here’s our guide to developing AI with the Secure AI Framework (SAIF) on Google Cloud.
The truths about AI hacking that every CISO needs to know (Q&A): How will AI boost threat actors? And what can chief information security officers do about it? Google’s Heather Adkins, vice-president, Security Engineering, explores how securing the enterprise is about to change.

What’s new in AI infrastructure and orchestration in October

Fri, 30 Oct 2026 16:00:00 +0000

The Google Cloud AI Infrastructure team had another busy month in October, filled with new product launches and updates to power today’s AI and agentic workloads scalably and efficiently. We were also honored to share our progress with customers, partners, and the broader community at leading industry events, including NVIDIA GTC, PyTorch Conference, and AGNTCon + MCPCon. Let’s dive in!

Product, technology, and tools updates

New compute service: Agent Compute Engine is a purpose-built compute service providing high-performance, hardware-isolated microVM sandboxes alongside expanded agent governance capabilities like Agent Gateway and Agent Registry, all in the service of securing and scaling autonomous AI agents. Its primary benefits include rapid sub-second activation, extreme workload density through copy-on-write memory sharing, enterprise-grade security and identity management, and cross-cloud observability.
AI Hypercomputer updates: At NVIDIA GTC Berlin 2026 we announced updates to our agentic infrastructure, including to A5X instances powered by the NVIDIA Vera Rubin NVL72 platform, Virgo Network scale-out fabric, and Managed Lustre storage. These co-engineered advancements provide enterprises with accelerated performance, enhanced sovereign AI controls, while advanced Confidential Computing capabilities help you train and deploy complex multi-step models securely and efficiently.
GKE product update: CPU startup boost for GKE, in preview, uses Kubernetes In-place Pod Resize to dynamically elevate container CPU resources during initialization and scale them back down to steady-state levels without having to restart the containers. This feature removes startup throttling and accelerates cold starts by up to 2x for demanding workloads, helping to lower cloud infrastructure costs.
GKE product update: The preview of the Wiz Runtime Sensor add-on for Google Kubernetes Engine (GKE) brings low-touch deployment, automated lifecycle management, enterprise credential security, and fleet-wide governance to Wiz's eBPF runtime threat detection for both Standard and Autopilot GKE clusters.
Modernization tools: Together at last. Google Cloud Modernize brings Google Cloud’s migration and modernization tools, including Migration Center, Google Cloud VMware Engine, Google Cloud Mainframe Modernization and our new EKS-to-GKE Migration Agent, into a single portfolio. Enterprise teams now have purpose-built agentic capabilities across infrastructure assessment, platform modernization, and application modernization.
Cloud Run product update: Temporal Serverless Workers on Cloud Run automatically scale worker pools based on Temporal task queue demands and scale down to zero when idle. This integration eliminates cluster management, lowers costs, using Temporal Worker Versioning to safely handle stateful, long-running workflows and AI agent workloads.
New open-source product: llm-d 1.0 is an open-source, Kubernetes-native serving stack that turns a cluster of model servers into a coordinated serving pool, unifying distributed LLM inference across mixed Google Cloud TPU and GPU fleets. Co-founded by Google and partners, it provides production-ready benefits such as cache-aware routing, tiered KV cache offloading, and priority-based flow control to dramatically improve context reuse, reduce time-to-first-token latency, and efficiently scale multi-tenant agent workloads.
New open-source product: Working with Meta, we open-sourced TorchTPU, a native PyTorch stack that integrates Google TPUs directly into the PyTorch ecosystem alongside GPUs and CPUs. Now developers can train, serve, and deploy models natively on TPUs using familiar frameworks like Hugging Face, vLLM, and TorchTitan.
Google Distributed Cloud updates: Organizations that operate under strict sovereignty mandates would love to use AI to help them process data, but can’t send it to the cloud. The latest version of Google Distributed Cloud (GDC) air-gapped brings frontier AI models and multimodal enterprise search with Gemini into these isolated environments. In addition, GDC air-gapped supports NVIDIA HGX B300 systems based on the NVIDIA Blackwell Ultra architecture, and features enhanced networking and security capabilities.
Edge inference product: Google open-sourced ML Drift, a cross-platform on-device GPU compute engine built for LiteRT that abstracts hardware complexities across OpenGL ES, OpenCL, Metal, and WebGPU. The framework delivers peak performance, enhanced 5D tensor support, and memory efficiency for both classical machine learning workloads and advanced generative AI models at the edge.
Open-weight model support: Reflection's Beam is an open-weight, 501-parameter model released under an Apache 2.0 license designed specifically for frontier software engineering and long-horizon agentic tasks. We announced day 0 support and ready-to-run deployment recipes for service Beam on AI Hypercomputer and NVIDIA AI Infrastructure.

Practitioner guides and how-tos

AI network reference architectures: In Networking for AI inference model serving - GKE only and for all other backends, learn about reference architectures for AI inference model serving that provide centralized governance and private IP routing using Private Service Connect, Apigee API Management, and Model Armor. These architectures simplify multi-model scaling and prompt security across GKE, Cloud Run, and hybrid clouds.
How-to guide: A Tale of TPU Observability on GKE: Part 2 - Dynamic Slicing Enhancements breaks down GKE's new Dynamic Slicing observability tools and metrics for next-generation v7 TPUs through an engaging, story-driven observability guide featuring Alex, a lead AI platform engineer. Explore the step-by-step diagnostic workflows with real-world examples and handy command-line snippets. Check out Part 1 here.

Customer and partner updates

AI21 achieves an 83% reduction in time-to-start for AI workloads with AI Hypercomputer: Frontier lab AI21 needed a new stack on which to train its Jamba family, and chose AI Hypercomputer configured with GKE on A3 and A3 Ultra instances. By doing so, AI21 says it cut high-priority job wait times from 72 hours to 12, and manual scheduling interventions from 20 per week to zero.
Visko streams real-time world models on Google Cloud TPUs: Visko, an AI startup developing interactive world models, deployed an end-to-end live inference pipeline that combines video generation, super-resolution upscaling, and synchronized audio onto a single Cloud TPU slice. In this blog, the team discusses how it engineered the pipeline, the architectural optimizations it made, and practical lessons for other startups building ultra-low-latency generative AI.
ZenteiQ scales physical AI inference 4x with GKE and llm-d: In order to pre-train models that understand physical laws, governing equations, and numerical solvers, Indian AI startup ZenteiQ designed and built a large-scale batch synthetic data-generation pipeline that runs across multiple model families and feeds verified datasets into training runs. GKE and GKE Inference Gateway gave ZenteiQ the inference-routing layer within its pipeline to scale aggregate inference throughput 4.07x and grow daily output from 1.47 billion to 5.97 billion generated tokens for every 1,000 accelerators.
Palo Alto Networks modernizes cloud-delivered security service: Palo Alto has worked with Google Cloud since 2018 to deliver its Prisma Access security service, and it was time to update it. In order not to compromise its security guarantees, Palo Alto separated the platform into two independently scalable layers: an ingress layer that handles traffic routing and forwarding running on GKE, with Axion C4A instances; and a security layer that handles threat inspection, running on dedicated single-tenant VMs fronted by an Internal Load Balancer. Each layer can now scale, update, and operate independently.
Oracle is the first hyperscaler to provide TPUs outside Google Cloud: By deploying Google TPUs, Oracle is broadening the range of AI chips available through its Oracle Cloud Infrastructure (OCI), and giving customers more options for running AI workloads. Google TPUs on OCI will be available in 2027.

Research, reports and deep-dives

New resource: As enterprise organizations shift from isolated AI experiments to production-scale systems, different departments — from engineering and finance to executive leadership — often speak different technical dialects; The Encyclopedia of AI Infrastructure bridges that communication gap by demystifying complex infrastructure jargon.
Analyst report: We are thrilled to announce that Google was named a Leader in the inaugural 2026 IDC MarketScape for Agentic AI Platforms.
Deep dive: A companion technical guide describing the architecture of Agent Compute Engine and how Google’s Gemini team needed both the isolation of a VM and the speed and density of a container to deliver personal agents for every Gemini user.

September 2026

Product, technology, and tools updates

Google Kubernetes Engine updates: The GKE team is all about improving the scalability of the platform, and in September, those improvements came in many shapes and sizes:

New feature: Need an execution runtime with higher density for your agentic workloads? We engineered the new open-source GKE Agent Substrate to run millions of sandboxes with 10x higher density than standard container runtimes. Agent Substrate also delivers sub-500ms resume operations at over 500 suspend/resume activations per second with a native zero-trust kernel and network isolation.
Product update: GKE now has scale-to-zero capabilities built-in. No need to configure complex components to scale your workloads down, thanks to the HPA with the Autoscaling Metric and support for KEP-2021, which do the job for you, out of the box. Read the blog to learn more.
Product update: Further, the GKE HPA (with the above-mentioned Autoscaling Metric) now lets you scale up and down based on custom PromQL metrics, in addition to standard metrics, allowing you to trigger workloads according to conditions that are meaningful and unique to your business. Read more here.
New feature: Yet another scalability feature is GKE Pod snapshots, which lets you save the running state of your workload, including CPU and GPU memory, and restore it on demand. According to internal tests, GKE Pod snapshots can reduce AI inference start-up by as much as 89%. Learn more here.
New migration tool: Finally, if you’ve always wanted to migrate your container workloads from AWS EKS to GKE but feared a daunting, high-friction engineering endeavor, we’ve just launched GKE agentic migration, a purpose-built agent plugin that replaces brittle, ad-hoc prompting with an AI-assisted migration pipeline protected by deterministic guardrails. Designed as a compilation of agent skills and a local Model Context Protocol (MCP) server, it uses AI to translate complex AWS EKS IaC and Kubernetes manifests directly into GKE landing zones. Get started with the onboarding guide.
Feature updates: Reinforcement learning (RL) and evaluation workloads are a beast: In a standard agentic RL loop, an LLM policy generates actions like code snippets on GPUs and executes them inside isolated CPU sandboxes to observe a reward signal. However, when scaling up this loop to support tens of thousands of parallel rollouts, infrastructure bottlenecks emerge, for instance idle accelerators, image cardinality, and a saturated control plane. To help, we developed GKE Agent Sandbox optimized for RL, plus an Agent Sandbox RL orchestration SDK and native integrations for popular RL gyms and harnesses. All are now generally available, and you can learn more here.

Storage updates: AI trains and creates lots of data, and that data has to live somewhere — in block storage systems, file systems, object stores and databases. We announced enhancements to our storage portfolio to help this critical layer roll with the agentic punches:

Product update: Filestore agent volumes offer high-performance, elastic, persistent file storage for agentic workloads. Thanks to its tight integration with GKE Agent Substrate and GKE Agent Sandbox, Filestore agent volumes automatically allocates and attaches a dedicated, isolated file workspace to GKE agent sandboxes in milliseconds. Request access to the preview here.
New product: If you run generative AI and RAG data layers — think Milvus, Pinecone, Qdrant, Vespa, Redis, and in-memory context caching — you may want to take a look at the M4N family of VMs, now GA, which offers the highest per-core IOPS and throughput of leading hyperscalers. Paired with Google Cloud's custom Titanium offload architecture and paired with Hyperdisk Extreme, M4N instances deliver up to 25,000 MiB/s (25 GiB/s) of aggregate host storage performance and up to 1 million IOPS.
New product: Another new Compute Engine product, Z4D, is GA, and a strong storage solution for AI/ML training and inference workloads. When configured as a bare metal instance, Z4D provides both the high local SSD (LSSD) capacity and low latency required by agentic microVMs, so you can run thousands of isolated sandboxes per host with native performance and efficiency. Learn about Z4D machines here.
Product update: Today’s AI training and inference pipelines create data faster than most storage management systems can keep up, creating challenges for teams trying to understand their storage estates. A new version of Storage Intelligence advisor makes it easier to answer the question: "What’s in my buckets?" and quickly identify unexpected changes. Then, enhanced batch operations let you automate bulk changes across your buckets — say, move storage classes, mass-delete stale or temporary data, or apply metadata, tagging, retention, or encryption changes. Learn about the latest in Storage Intelligence advisor here.
New product: Last but not least, a new version of AlloyDB for PostgreSQL brings together pioneering Google infrastructure — Colossus distributed file system, and Jupiter network — to power a new, no-compromises database architecture for the agentic era.

Practitioner guides and how-tos

How-to: Wish you could make GPUs and TPUs scattered around the globe behave as a single pool behind a single entry point? In this blog, we show you how to do just that. At the edge, the multi-cluster GKE Inference Gateway focuses on global, multi-region traffic distribution and high availability. Beneath that, the LLM-d router handles complex, memory-aware scheduling algorithms to keep utilization high. This architecture is deliberately runtime-, model-, and accelerator-agnostic, and in tests, routing traffic through the multi-cluster GKE Inference Gateway added less than 1% overhead.
Guide: Using or planning to use GKE on TPUs for AI model training or inference? Training massive Large Language Models (LLMs) or running high-throughput inference serving represents a significant investment in specialized AI hardware, such as Cloud TPUs and GPUs. To get the most out of every dollar spent, you need to understand workload lifecycle metrics in GKE. This detailed guide explains how to turn opaque cluster behaviors into actionable telemetry.

Customer and partner updates

GKE customer: Learn why gaming startup SeaVerse relies on GKE Agent Sandbox for its multi-tenant workloads, and how the platform helped it decrease its infrastructure costs by 60%.

Research, reports and deep-dives

In case you missed it, we’re also thrilled to share that Google has been named a leader, including achieving the highest score on either product or strategy, in three key analyst reports from Gartner and Forrester.

Google is a leader in The Forrester Wave™: Public Cloud Platforms, Q3 2026: Highest overall score of any cloud provider!
Google named a Leader in 2026 Gartner® Magic Quadrant™ for Strategic Cloud Platform Services: Positioned furthest for “Completeness of Vision” of all vendors evaluated.
Google is a Leader in the 2026 Gartner Magic Quadrant for Container Management: Positioned highest in “Ability to Execute” of all vendors evaluated.
Google Cloud achieved a Gold rating in the latest SemiAnalysis ClusterMax 3.0 report, which evaluates the reliability, performance, support, pricing, and security of GPU providers globally. See the full report for more.

August 2026

Product, technology, and tools updates

Product update: Filestore, Google Cloud’s first-party, secure, scalable NFS file service, has emerged as a popular storage platform for AI and agentic workflows, and now, it’s even better suited to the task, with a new backend storage layer built directly on Colossus, Google’s foundational distributed storage system. This new backend lets you provision IOPS independently from storage capacity, and is deeply integrated with GKE. In AI environments, this can help you service so-called agentic swarms — large groups of agents that need to read and write to a common dataset — without a drop off in performance. For more, check out the blog post.
New feature: gVisor sandboxes are now available in distributed Ray clusters on GKE. In partnership with Anyscale, we introduced an experimental library for Ray that brings gVisor, Google’s open-source application kernel, directly into distributed Ray clusters. gVisor provides lightweight environments with stronger isolation than ordinary containers, plus fast startup times and low memory overhead. To try out these sandboxing capabilities on GKE, head over to the Ray sandboxing User Guide.
Product update: Looking for high-performance, easy-to-use infrastructure on which to run a personal AI agent, but don’t want to spend a lot of money? New Cloud Run instances are dedicated, singleton compute runtimes on Cloud Run that won’t shut down when the agent is idle. Better yet, the cost to run a Cloud Run instance with 1 vCPU and 1 GiB of memory continuously for 30 days is just $5.70.

Practitioner guides, documentation and how-tos

How-to guide: Big news in Model Context Protocol (MCP) land: As of the 2026-07-28 specification, the protocol core is “completely stateless. The handshake is gone. The initialize / initialized handshake (SEP-2575) and the logical Mcp-Session-Id header (SEP-2567) have been removed entirely. Instead, every request is now self-describing and independent.” Whoa. Learn more about the changes that the latest MCP specification brings, and more importantly, how to implement them, in this Google Developers blog.
Guide: Real-time AI systems make a mess of traditional network load balancing techniques. “Instead of handling isolated requests, the backend has to manage a continuous, live bidirectional stream. You’re dealing with a constant stream of audio chunks, transcripts, model outputs, and synthesized speech flowing back and forth simultaneously.” Things only get worse when the user gets involved. “The server has to immediately halt its current speech generation, pivot to update the context, maybe trigger a new tool, and start drafting a different response; this must be done without dropping the connection.” For a new approach to managing load in the AI era, read Scaling real-time AI agents with session-aware load balancing.
How-to: Learn how to build an elastic, scalable LLM inference platform on GKE, even with a mix of different GPU accelerators. The proposed architecture combines Capacity Advisor and Compute Advisor, plus high-performance storage like RunAI:model streamer or GCPFuse with parallel downloads. Get all the details here.
Documentation: The thing about hosts with GPUs or TPUs is that you can’t use live migration to update them, setting up a maintenance challenge. In this new docs page, learn how to update accelerator-equipped hosts according to your tolerance for downtime for your training and inference workloads.
Documentation: Advanced Compute Images, or ACIs, are standardized image stacks for AI/ML and HPC infrastructure, so you don’t need to manually build your own custom images. In this new docs page, learn how to create an ACI image using the Google Cloud CLI, console, or SchedMD's Slurm workload manager.
Guide: AI workloads are notoriously difficult to architect, resource-intensive, and bursty, which can also lead to scaling bottlenecks and large pools of underutilized — or misutilized — compute resources. A new blog outlines the three main ways to achieve dynamic capacity management in Google Cloud: 1) scheduling capacity for planned downtime; 2) maintaining automated fallback capacity for unplanned downtime; and 3) relying on GKE’s core orchestration capabilities to automate resource allocation.

Customer and partner updates

Business orchestration software provider UiPath was dealing with spiky workloads, and wanted more predictable costs. To get there, it re-architected its infrastructure, moving from isolated clusters to a shared Google Cloud GPU fleet that included both A3 VM instances (NVIDIA H100 GPUs) for training with G4 VM instances (NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs) for inference. You can read more about their architecture here.
Mirendil, an frontier AI lab focused on accelerating AI development, announced that it is using AI Hypercomputer with both TPUs and NVIDIA GPUs to support its model pre-training and post-training applications.
Replenit, a retail CRM provider, built its AI decision engine in Google Cloud, using BigQuery, Gemini Enterprise Agent Platform, and open-source Gemma models that it runs on Cloud TPUs. This latter combination provided Replenit with 90% lower pipeline costs than their previous cloud provider, the company reports. Read the full case study for more.
Malachyte architected its AI-powered e-commerce recommendation platform on top of Bigtable, Managed Service for Apache Kafka, Pub/Sub, Compute Engine, and last but not least, GKE. See how it all comes together in this blog.

July 2026

Product, technology, and tools updates

Product update: Google Cloud Managed Lustre is now GA, and available in four distinct performance tiers that deliver throughput ranging from 125 MB/s, 250 MB/s, 500 MB/s, to 1000 MB/s per TiB of capacity — with the ability to scale up to 8 PB of storage capacity. The Managed Lustre solution is powered by DDN’s EXAScaler, combining DDN's decades of leadership in high-performance storage with Google Cloud's expertise in cloud infrastructure.
Product update: C4N network and storage optimized VMs are now GA. C4N is our first network- and block-storage-optimized VM series built to eliminate data-transfer bottlenecks. Powered by 5th Gen Intel Xeon Scalable processors and built on Google's Titanium offloading hardware, it achieves 400 Gbps network bandwidth, 95 million packets per second (MPPS), and up to 25 GiB/s of block storage throughput when paired with Hyperdisk Extreme.
New feature: GKE Dataplane V2 up to 15K Nodes with Network Policies (GA). This capability enables standard GKE clusters to scale up to 15,000 nodes while maintaining full active Network Policy enforcement, supporting the massive infrastructure needs of large enterprise and AI/ML customers.
New feature: Co-operative time-slicing in llm-d. If you’re running reinforcement learning (RL) workloads, you can now interleave independent RL jobs onto shared physical hardware, increasing aggregate accelerator duty cycles from a ~40% baseline up to 70% without impacting model convergence or accuracy.
New AI security tool: Looking to secure your AI supply chain on GKE, deploy AI workloads safely, and cut down on shadow AI? We open-sourced k8s-aibom, a lightweight, unprivileged Kubernetes controller that continuously monitors container clusters to automatically detect running AI runtimes (like vLLM and Triton) and generate standard CycloneDX Machine Learning Bill of Materials (ML-BOMs). Check out the k8s-aibom project and get involved.

Practitioner guides and how-tos

How-to guide: On July 27, Google announced Day 0 support for Moonshot AI’s Kimi K3 2.8-trillion-parameter open-weight model, the day weights were released. Whichever your preferred deployment path — via Model Garden, custom orchestration, or GKE with llm-d recipes — this guide offers detailed step-by-step instructions to help you evaluate and pilot Kimi K3 in Google Cloud.
How-to guide: Google Kubernetes Engine (GKE) managed DRANET supports both GPUs and TPUs. There are several configurations to use this implementation, including standard cluster (where you have full control) and autopilot cluster (where Google does the heavy configs for you). Take a deeper dive in the hands-on lab, GKE Autopilot clusters with TPUs, GKE managed DRANET and Gemma 4.
How-to guide: Learn to run Ray on TPUs, not GPUs. In Part 1 of this two-part series, we discuss TPU slices (hint: Ray thinks of them as just another accelerator on which to schedule), then walk through Ray’s various AI libraries (Part 2).
How-to guide: Evaluate TPUs for sample workloads using a new microbenchmark suite that helps you accurately assess whether a device is achieving its theoretical performance specifications, and to identify specific performance gaps or architecture-specific bottlenecks. Dive in here.
How-to guide: Scale your agents without killing your budget. Learn how GKE orchestration can help you safely pack more agents onto a fixed compute footprint with GKE Agent Sandbox and Pod snapshots. Whether your goal is performance or cost optimization, we teach you how to turn the right dials for optimal agent efficiency.
Technical blueprint: Inside the optimization of Mistral 3 large inference on Ironwood. This blog outlines how one Google team optimized Mistral 3 large MoE model inference on Google’s Ironwood (TPU v7x), achieving a 1.5x performance gain. They did so with hybrid sharding, replacing linear VPU summations with tree reductions, optimizing GMM/MLA kernels, and adopting asynchronous scheduling. As a result, they boosted throughput by up to 48% while maintaining benchmark accuracy neutrality. Read the full blog here.

Research, reports and deep-dives

Report: Google was named a Leader in the inaugural GartnerⓇ Magic Quadrant™ for AI Infrastructure, positioned highest for ‘Ability to Execute’ and furthest for ‘Completeness of Vision’. Gartner called out Google’s proprietary scalable compute, integrated AI Hypercomputer architecture, and the scale of our AI compute capacity as key strengths. Download a copy here.
Report: We recently surveyed more than 1,400 senior IT leaders for our State of AI Infrastructure report, and a resounding pattern emerged: The gap between AI ambition and infrastructure reality is widening. In fact, 83% of organizations say they require infrastructure upgrades to support production-grade agentic AI. Read the accompanying blog to understand how adapting your infrastructure to meet the demands that agentic applications place on your systems will help you move from pilot to production.

June 2026

Product, technology and tool updates

Product update: Protecting sensitive data used with AI is a critical part of advanced and secure cloud infrastructure. Confidential Computing cryptographically protects data in use in hardware-based Trusted Execution Environments (TEEs) with verifiable data integrity, and is now available on the accelerator-optimized G4 machine series, featuring NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. Get started with Confidential G4 VMs and Confidential G4 GKE Nodes.
Developer resource: The new TPU Developer Hub is the place to go for model builders, optimizers, and developers to learn to unlock the full performance of Google Cloud TPUs. Read more in this blog.
New product: Scale your AI workloads with the new OpenTelemetry-Based TPU AI Telemetry Collector Agent. For the first time, you can route high-fidelity TPU hardware telemetry to Google Cloud Monitoring, Google Managed Prometheus, or your own self-hosted Grafana stack.

Practitioner guides and how-tos

How-to guide: Learn how to build high availability into an AI inference workload running on GKE Inference Gateway with TPUs, Cloud Storage FUSE and Dynamic Resource Allocation (DRA). This blog provides an overview, or you can get all the technical details in the hands-on codelab.
How-to guide: Did you know you can connect your AI agents to unstructured data in Cloud Storage via Model Context Protocol (MCP)? In this blog, learn about why would want to do that from three customer examples, then how to do it, choosing either a fully managed service, or a self-managed local server for more customization and control.

Research, reports and deep-dives

Report: According to an independent benchmark report, GKE Inference Gateway outperforms the next leading managed Kubernetes service with 15.7% higher throughput, 92.8% shorter wait times, and 62.6% lower inter-token latency. This performance can be attributed to its use of prefix caching, which optimizes LLM performance by storing the KV cache (activation states) of long, repetitive prompt prefixes. Learn more in the blog.
Architecture deep dive: A closer look at the cold start problem, this time for TPUs and GKE, and how the Run:ai Model Streamer can help change the dynamic.

Customer and partner updates

Customer win: Leveraging GKE, BigQuery, Cloud SQL, and Gemini Enterprise Agent Platform, Pager Health is eliminating operational fragmentation to deliver a simplified, personalized U.S. healthcare experience that transforms lives.
Customer win: Trustpilot, the customer review platform, built a high-volume streaming pipeline using fine-tuned Gemma models with Dataflow and Gemini Enterprise Agent Platform running on cost-optimized A2 VMs using A100 GPUs, as well as optimized version of vLLM maintained by Gemini Enterprise Agent Platform.

May 2026

Product, technology and tool updates

Product update: GKE Agent Sandbox is now generally available.
New open-source project: Agent Substrate is a new open-source project aimed at continuing to push the limits of agentic infrastructure density
New feature: Google AI Edge Portal, a solution for testing and benchmarking on-device machine learning (ML) at scale, now supports benchmarking and debugging on-device LLMs. Read more here.
Product deep dive: We went into depth about Cloud Storage Rapid, a new family of high-performance storage offerings for AI workloads. At launch, offerings include Rapid Bucket (formerly Rapid Storage), a high-performance zonal object storage offering, and Rapid Cache (formerly Anywhere Cache), which accelerates reads on-demand and colocates compute and data for workloads in existing buckets.

Research, reports and deep dives

Architecture deep dive: Google Global Infrastructure VP Bikash Koley and Engineering Fellow Arjun Singh provide a high-level overview of the challenges that AI workloads pose to network infrastructure, and discuss the deep enhancements we’ve made to our data center fabrics, WAN, and global networks to better support them.
Architecture deep dive: We unveiled a new cluster-level reliability model for developing frontier AI models on TPUs, ditching instance-level reliability

Customer and partner updates

Customer win: Visual media provider Imgix serves more than 8 billion images and videos from AI Hypercomputer equipped with G4 VMs powered by NVIDIA RTX PRO 6000 Blackwell GPUs.

Innovate on your own business terms: Google Cloud spend caps now generally available

Fri, 30 Oct 2026 16:00:00 +0000

Today, we’re announcing that spend caps on Google Cloud Budgets are generally available.

With this release, we’re delivering ways to protect your workloads from runaway costs without risking resource deletion. More importantly, we’re introducing three major updates, directly inspired by your feedback since the initial preview, to bring programmability, real-time visibility, and automated guardrails to your cost-control playbook.

How the new spend caps can help your business

We have significantly enhanced the usability and power of spend caps in three ways:

1. Automation through spend caps API: Managing guardrails manually across hundreds of projects is a major operational challenge. To address this, we have launched the spend caps API, which can be accessed from the existing Cloud Billing Budgets API. You can now programmatically enforce your financial boundaries for eligible services within each project, without additional overhead. You can automatically lift, update, or delete caps on the fly and instantly apply a spend cap whenever a developer spins up a new sandbox project.

2. Real-time visibility with live cost reports: Spend caps for AI services trigger within minutes of usage using estimated, live costs. When a cap triggers, you need to know exactly why and when the limit was breached. You can now view these faster, estimated costs directly inside your Cost Reports. Instead of waiting hours for standard billing pipelines to settle net costs, you get the exact financial snapshots that triggered the enforcement, giving you the immediate context needed to run a precise root-cause analysis.

3. Automated guardrails for your monthly resets: Before, once a spend cap was triggered, the usage block remained locked until you manually lifted it in the UI. Today, we are introducing automated monthly resets. You no longer need to worry about manually lifting caps every month. Enforced spend caps will automatically reset, and blocked services will seamlessly resume at the beginning of the next month. However, if your governance policies require a strict manual audit before resuming operations, the manual override option continues to exist exactly as before.

Best practices for deploying spend caps

As you roll out spend caps across your organization, keep these operational tips in mind:

Target non-production environments first: Start by applying strict, low-dollar spend caps to your developer sandboxes, testing clusters, and experimental AI agent playgrounds.
Configure alerts wisely: Default alerts are automatically sent to billing administrators and project owners at 50%, 80%, and 100% of the cap. Make sure the right folks are set up in these roles. This ensures your team is never caught off guard when a budget is nearing depletion. Spend caps permissions can also be added to custom roles, if needed.
Understand fixed vs on-demand costs: Spend caps halt pay-as-you-go and CUD-covered requests once the budget threshold is hit. However, Provisioned Throughput is never halted; requests using reserved capacity will continue to run, and serve traffic normally.

Refresher on GCP’s dual-pronged defense playbook

Spend Caps work seamlessly alongside our native Early Anomalies Detection to provide a closed-loop defense for your cloud spend. Read our previous blog for a detailed explanation on how to use these tools to effortlessly set up a GCP defense playbook.

Playbook step	Feature	Focus	Core action	Key benefit
Step 1	Early anomalies	Predictive monitoring	Analyzes early cost signals daily to alert you of directional spend variances	Spot anomalous cost spikes early and pinpoint the exact SKU driving the trend
Step 2	Spend caps	Financial enforcement	Programmatically pauses usage for a specific project/service once a budget is breached	Absolute protection against runaway spend without deleting resources

Get started today

Spend caps are now ready to support your workloads at scale. Read our documentation for more details and navigate to the Budgets and Alerts page in your Google Cloud Billing Console, to set your first spend cap.

To integrate programmatic guardrails into your deployment workflows, check out our newly published spend caps API documentation.

Scale agents securely with IAM Helper Agent Skills and MCP Tools

Tue, 27 Oct 2026 16:00:00 +0000

As organizations adopt autonomous agents to build applications, securing identities and access has become more challenging. Developers relying on agents to provision cloud resources, write Terraform, and deploy microservices at machine-speed only find their CI/CD pipelines blocked even faster as security friction scales.

To help you get ahead of agentic-era Identity and Access Management, today we’re introducing IAM Helper. This suite of intelligent, agent-facing building blocks can automate complex, end-to-end IAM workflows.

As an agent-native orchestration and interface layer for Google Cloud IAM, it augments Google Cloud’s core IAM engines — including Policy Intelligence and Privileged Access Manager (PAM) — making IAM skills callable directly by autonomous code assistants, deployment pipelines, and custom agents. Using this functionality reduces the administrative overhead of maintaining least-privilege security and agent governance, while maximizing utility, and accelerating deployment.

IAM Helper Agent Skills: Fast, secure deployment with automated governance

Manual troubleshooting of permission blockers creates friction for engineering and administrators, while shortcuts can lead to broad agent over-permissioning that puts least privilege at risk.

IAM Helper empowers agents to handle access management dynamically. It shifts the burden of policy evaluation away from human administrators and LLMs, offloading complex orchestration directly to the Google Cloud backend, to maximize token efficiency and prevent costly operational bottlenecks.

Rather than operating as isolated utilities, IAM Helper Agent Skills guide agents through the cloud access lifecycle. These end-to-end workflows, helper scripts, and guardrails packaged using the open Agent Skills standard can orchestrate multiple underlying functions and APIs to resolve complex IAM tasks. They work across common agents (Gemini CLI, Gemini Cloud Assist, ADK, Claude Code, and Cursor).

Built-in guardrails and real-world impact

Every IAM Helper agent skill separates read-only discovery from policy changes: Diagnostics and simulations run automatically, while any change to access policies pauses for human-in-the-loop (HITL) confirmation based on customizable risk tiers. To see the value of IAM Helper in practice, consider how it transforms two everyday scenarios:

Scaled user-assisted remediation: Suppose a developer's coding agent hits a 403 error while deploying a new application. Rather than filing a support ticket and idling, the agent uses iam-helper-for-troubleshooting to identify the missing Pub/Sub permission in seconds, then uses iam-helper-for-privileged-access-management to request a one-hour JIT grant, eliminating hours of debugging friction.
Autonomous pipeline self-healing: An infrastructure agent encounters a permission block during a routine deployment pipeline. Rather than failing the build, the pipeline uses IAM Helper Agent skills to generate a revised, least-privilege policy, verifies against 90 days of logs via iam-helper-for-policy-simulator that active services won't break, and routes the change to an admin for quick approval.

IAM Helper Agent Skills are designed for maximum token and cost efficiency, allowing agents to execute multi-step administrative operations autonomously.

Authoring compliant, least-privilege policies: As DevOps engineers translate access requirements into JSON, YAML, or Terraform HCL, iam-helper-for-policy-management generates and validates IAM v1 allow bindings and IAM v2 deny rules. It automatically catches and blocks risky configs before they are applied, including public access (allUsers) and broad basic roles (roles/owner, roles/editor).
Unblocking 403 errors in seconds: When an on-call engineer or deployment agent hits a cryptic 403 PERMISSION_DENIED error, iam-helper-for-troubleshooting identifies the exact missing permission or blocking rule in seconds. It adapts to the user’s role: Developers can check if they can unblock themselves locally or generate a structured ticket for their administrator (Requester flow), while security administrators can trace inherited allow and v2 deny rules across organization, folder, and project hierarchies to resolve the issue (Resolver flow).
Simulating blast radius before rolling out changes: Before a platform engineer updates a policy or applies IAM Recommender role reductions, they need confidence that active workloads won't break. iam-helper-for-policy-simulator walks the agent through a structured plan, simulate, analyze, and apply flow, testing proposed changes against 90 days of historical access logs to confirm existing traffic won't be disrupted before any update takes effect.
Elevating access just-in-time instead of granting standing privileges: When responders or maintenance agents need temporary production access, iam-helper-for-privileged-access-management manages the lifecycle of on-demand credentials through Privileged Access Manager (PAM). It supports requester, administrator, and approver workflows so elevated access is audited, time-bound, and automatically revoked once the task is done.

While IAM Helper Agent Skills supply the high-level workflow logic, they are all based on a common set of foundational MCP Tools.

Powered by Google Cloud IAM MCP tools

Google Cloud IAM Helper MCP tools provide the backend connection through atomic, read-only functions exposed via Google-managed MCP servers across core IAM endpoints (Policy Troubleshooter, Policy Analyzer, Policy Simulator, and IAM v1/v2 APIs).

Rather than requiring teams to write custom API wrappers, MCP tools give agents a standard way to query the cloud backend and get exact answers without parsing verbose policy documents in the prompt, maximizing token efficiency. Agents can dynamically discover and invoke these tools to execute basic operations like checking permissions or listing roles when custom logic is required.

Because operations like hierarchy traversal, group expansion, and policy evaluation happen on dedicated IAM endpoints (spanning IAM v1 allow roles, IAM v2 deny, Policy Troubleshooter, and Policy Analyzer), skills run quickly and accurately with minimal token usage. Teams can also use these atomic MCP tools directly when writing their own custom Agent Skills, making it straightforward to combine Google Cloud's IAM capabilities with internal ticketing systems, custom compliance checks, and existing deployment pipelines.

Getting started

Accelerate your agentic workflows, reduce operational overhead, and bring intelligent, scalable security to your deployment pipelines. Start building with IAM Helper Agent skills today by exploring our Google Cloud Agent Skills Catalog and dive deeper with the official IAM documentation.

Gemini Robotics Embodied Reasoning 2 is now available on Google Cloud

Tue, 27 Oct 2026 13:00:00 +0000

Earlier this year, we introduced Gemini Robotics 2 — our next-generation intelligence layer enabling robots to think, act, and adapt in the physical world. At the core of this breakthrough is Gemini Robotics Embodied Reasoning 2 (ER 2), a specialized Gemini model that serves as a robot’s high-level brain. Working across any robot embodiment — from stationary arms to full humanoids — Gemini Robotics ER 2 combines advanced spatial and temporal video understanding with long-horizon planning and built-in safety guardrails. This enables robots to converse naturally with humans, orchestrate low-level action models and hardware APIs, execute multi-minute tasks involving hundreds of decisions, and collaborate across multi-robot teams.

Today, we are bringing this same frontier embodied reasoning model to enterprise production environments: Gemini Robotics ER 2 is now available in public preview on the Gemini platform. For the first time, organizations can pair Google DeepMind’s latest embodied reasoning model for robotics with the enterprise-grade security, multi-agent orchestration, and end-to-end lifecycle management of Google Cloud — making it easier to simulate, evaluate, and deploy safe, adaptive physical AI at scale.

Powering physical agents for enterprise robotics

Gemini Robotics ER 2 transforms a robot into a physical agent. The model combines deep spatio-temporal reasoning and long-horizon planning, enabling robots to orchestrate steps to complete a task, self-correct, and generalize to novel situations. To build an agentic setup, developers can declare low-level control interfaces — like Vision-Language-Action (VLA) models or navigation APIs — as tools, and stream multimodal video, audio, or text directly into the model.

Here is what the model delivers for businesses building and operating robotics:

1. Physical task verification, understanding, and self-correction

In physical operations, a robot needs to know whether a step actually succeeded before moving on to the next one. Gemini Robotics ER 2 adapts to new and unfamiliar situations on the fly by reasoning over continuous video feeds. The model:

Classifies task progress into binary status, discrete categories, or continuous scores to improve real-time situational awareness and enable real-time adjustment or iteration on failed steps, without restarting an entire workflow.
Pinpoints the exact video frame where a critical event occurs (such as when to stop pouring a liquid) with 91.3% accuracy and a 0.96-second mean absolute frame distance, running at 4x the execution speed and a fraction of the compute cost of many larger model categories.
It advances our core spatial reasoning capability, including visual question answering and instrument reading. For example, it can read and interpret values across 10+ instrument categories including circular dials, sight glasses, digital displays, linear scales, rulers, and liquid thermometers directly from visual input.

2. One intelligence layer across any embodiments

Gemini Robotics ER 2 orchestrates robots of many shapes and sizes, scaling its intelligence from stationary arms and wheeled rovers to quadrupeds and complex forms. It also unlocks multi-robot collaboration, allowing different types of robots to share a semantic understanding of their workspace, divide labor, and hand off tasks to complete workflows a single machine could not do alone.

3. Agentic tool orchestration and natural language interaction

Gemini Robotics ER 2 thinks about what comes next while simultaneously orchestrating actions. Developers can declare low-level control interfaces — such as VLA models, navigation APIs, or custom functions — as tools across simulation, real-world robot control, or human tele-operation. Gemini Robotics ER 2 can orchestrate navigation and manipulator movement APIs from natural language commands, allowing the robot to explain its approach while performing an action, and letting operators redirect it without using technical language.

4. Build safe, adaptive physical agents in Google Cloud

Developers and enterprise teams use the Gemini platform to simulate, orchestrate, and manage physical agents, which are then deployed to run directly on physical robots at the edge using Google Cloud's scalable infrastructure:

Simulate fast: Test robot behaviors before touching hardware by running simulations in familiar environments like MuJoCo-Warp and NVIDIA Isaac Sim. Then generate synthetic scenarios with multimodal models and auto-label workcell imagery on the Gemini platform.
Train smart: Use Gemini Robotics ER 2's spatial pointing, moment finding, and progress classification to curate and label demonstration videos in Cloud Storage and BigQuery. Then train custom Vision-Language-Action (VLA) models through the Gemini platform, GKE, or Cluster Director.
Operate easily: Deploy Gemini Robotics ER 2 workflows on the Gemini platform with enterprise IAM, security controls, and built-in function calling. Offload latency-sensitive local workloads like SLAM and path planning to factory floors with Google Distributed Cloud, and stream fleet telemetry into BigQuery. Explore the full architecture in Google Cloud Physical AI solutions.

How customers are building with Gemini Robotics ER 2 in Gemini platform

“We’ve seen exponential growth in customers using AIVI-Learning every month, and Gemini Robotics ER 2 brings them even more value. Video changes what an inspection can be. Spot isn’t just checking whether something looks right in a moment in time, the ability to apply AI models to video opens a whole new world of possibilities.”

Marco da Silva, VP and General Manager of Spot at Boston Dynamics

“DeepReach AI is the real-world skills engine for physical AI. Our hardware and AI platform capture how skilled people work all over the world, and turn it into data flywheel for model training. Gemini Robotics ER 2 gives us accurate task understanding of every recording, to help us find data that matters most — faster and at a far greater scale.”

Chris Liu, Co-founder and CTO at DeepReach

“Advanced reasoning models have the potential to help robots interpret complex environments, plan actions and adapt to changing conditions, opening new possibilities for physical AI. Google’s Gemini Robotics ER 2 model represents an important step in this direction, and we’re excited to be working with Google on exploring how these emerging capabilities can shape the next generation of robotics compute platforms.”

Barrie Mullins, Senior Director of Arm Total Design for Physical AI, at Arm

“We aim to accelerate materials R&D workflows so our scientists can test more ideas, faster. Most lab automation today relies on rigid scripts for single, isolated tasks. By using Gemini Robotics ER 2, we are giving our robots the visual and physical flexibility to work with our existing infrastructure - whether that’s reading a touchscreen on a legacy microwave reactor that has no API, balancing and loading a centrifuge, or moving samples safely through a busy lab. This is the next level that we aspire to reach. This brings us much closer to self-driving labs that can run full experimental campaigns from synthesis to separation, freeing our researchers to focus on thinking up new ideas for discovering new materials.”

Prof. Loh Xian Jun, Executive Director at ASTAR Institute of Materials Research and Engineering (ASTAR IMRE) and Executive Director of MARVEL-AI”

Get started with Gemini Robotics ER 2 in Gemini platform

Developers and enterprise teams can start building with Gemini Robotics ER 2 today in preview via Model Garden in Gemini platform.

Build in Gemini platform: Try in Agent Studio. Visit the Gemini Robotics ER 2 documentation and pricing.
Explore the Developer Guides: Check out sample code and bidirectional streaming guides in the Gemini API Robotics documentation.
Read the Research: Dive into the benchmarks and safety technical report in the Google DeepMind Gemini Robotics ER 2 announcement.
Explore Physical AI on Google Cloud: Visit cloud.google.com/solutions/physical-ai and cloud.google.com/ai to learn how Google Cloud infrastructure supports simulation, training, and deployment.

Easy to integrate, compliant by design: Bringing the Swiss Franc to Google Cloud Universal Ledger

Mon, 26 Oct 2026 20:00:00 +0000

Every day, countless financial transactions are interrupted by slow, expensive, and inflexible conventional payment systems. As a result, public commercial stablecoins like Tether, USDC, and PayPal USD have become increasingly popular by making it easy to move money efficiently with instant, 24/7 blockchain settlement.

Yet major regulated banks have largely kept such transactions at arm’s length; for these financial institutions, reserve-backed public stablecoins present numerous risks, including strict capital and liquidity requirements and unmanageable regulatory and operational uncertainty.

Indeed, banks don't want to adopt existing stablecoins. Instead, they need the speed and efficiency of tokenized currency built on the trusted economic foundations they’ve nurtured over years or even decades in business. That’s why many are now turning to Google Cloud Universal Ledger (GCUL). Rather than channeling public stablecoins, GCUL provides the secure infrastructure banks need to issue and settle tokenized deposits. That brings real-time, 24/7 programmable payments into commercial bank money while satisfying regulators and without disrupting core banking operations.

For many banks, this is a compelling vision, but adopting new payment technology and integrating it with core banking systems can seem to bank CTOs like a daunting undertaking. That’s particularly true as they face mounting demands for the engineering resources that keep their businesses running while also striving to stay compliant with new regulations. But with a blockchain purpose-built for financial services like GCUL, this is anything but a herculean task.

Looking to expand its offerings for an increasingly global clientele, Switzerland’s Crypto Finance Group wanted to show it could deploy payment innovations like this in just weeks. That was its inspiration for launching a digital payments infrastructure pilot that enabled instant account-to-account transfers between customers of two Swiss banks on the GCUL network. The pilot, which was designed to operate within one of the world's most demanding financial security frameworks, provided a blueprint for how any commercial bank can build on Google Cloud technology to fundamentally modernize its payment operations.

Pain points in today's payments landscape

In recent years, there’s been a marked proliferation of tokenized payment methods. But stablecoins have transformed this previously niche technology, by effectively resolving some of traditional finance’s most enduring pain points, such as weekend transfer blackouts, direct transfers without correspondent banks, and fast settlement rails. Stablecoins also limit the frustrations caused by opaque cross-border transactions, where fragmented compliance checks can leave legitimate payments in extended and unexplained limbo and without a path to clear or timely resolution.

Still, commercial banks have been reluctant to offer popular reserve-backed stablecoins from permissionless blockchains to their customers despite their retail popularity. That’s because they come with numerous challenges to smooth and secure operations: an uncertain regulatory status, compliance issues around anonymous accounts, gas fees, and the scalability limitations of layer 1 blockchain technology. But perhaps banks’ biggest concerns about stablecoins is that they create a problematic economic disconnect — giving customers access to reserve-backed stablecoins reduces reserves and, as a recent McKinsey study found, banks’ ability to lend.

As a result, many banks are trying to figure out how to offer their customers the benefits of stablecoins — high speed, low cost, 24/7 availability, and programmability — without introducing problems often associated with existing reserve-backed stablecoins, such as regulatory issues, poor throughput performance, negative influence on bank reserves, and potentially no yield participation.

Solution proposal

Realizing this vision means modernized payments architectures must seamlessly and instantly enable individuals and institutions to transfer bank-held assets to accounts at any target bank. Orchestrating this shared ecosystem requires central entities to establish the governing rulebook, enforce issuance limits, and manage network-wide settlement parameters. During its pilot, Crypto Finance assumed this role, which within GCUL is known as the “currency operator.”

In this position, Crypto Finance can configure the network for either instant or deferred settlement. This choice is analogous to traditional banking: instant settlement acts like real-time gross settlement, where transactions settle individually and immediately and require high liquid reserves. In contrast, deferred settlement operates like netting, where payments are accumulated and offset and only periodically settle the net difference. For the pilot, Crypto Finance chose to adopt deferred settlement, in which GCUL manifests capital efficiency. In this way, clients enjoy instant transfers while banks conserve liquidity through periodic settlement.

Concept validation

To demonstrate instant, bi-directional account-to-account transfers on GCUL, Crypto Finance orchestrated a pilot with two digital token issuers, a Swiss digital asset bank, and a Swiss B2B transaction bank. Crypto Finance established the rulebook for interactions between the participant banks, such as interbank settlement, integration with traditional SIC payment rails, and AML/compliance protocols. Although some backend steps still needed manual intervention, Crypto Finance’s pilot proved that fully automated, cross-institution integration is feasible and established a generalized architecture that can scale to virtually any asset class.

Robustness and performance

A temporary pilot is one thing; moving into live client transactions carrying legal liability brings a whole new set of operational challenges. Live payment rails are considered critical infrastructure that demands security and availability. GCUL addresses these needs through redundancy between geographically distributed nodes operating on an enterprise-grade consensus model, ensuring reliability. Deploying client workloads on Google Cloud provides that enterprise-grade foundation and allows financial institutions to pair the highly resilient GCUL network with secure-by-design infrastructure, encryption at rest and in transit, and robust access controls that satisfy many of banks’ common regulatory, privacy, and data sovereignty mandates. While some banks may consider integrating their systems with fully public and permissionless blockchains, this approach introduces severe regulatory hurdles, anonymity risks, unpredictable gas fees, and low throughput. These are all roadblocks that in the past would have significantly delayed the pilot or rendered it a non-starter.

Ease of integration

The pilot's implementation, with systems integration completed in under three months, was compartmentalized across dedicated Google Cloud projects per participant. The currency operator, Crypto Finance, connected specialized components to a dedicated GCUL API. The first was a role admin service that controls account managers and token managers on the network. Next, it implemented a settlement registry to record when periodic off-ledger settlement between participant banks has occurred once a predefined threshold is reached. Finally, a cloud key management service (KMS) stored the currency operator keys to sign transactions.

At the same time, participant banks connected their core banking systems to their own GCUL API instances. But these participant banks don’t need a large Google Cloud footprint. Instead, they can connect to GCUL nodes via Remote Procedure Call from their existing on-premise deployments and manage private key material through solutions other than KMS. Regardless of the setup, transactions are seamlessly routed to the shared global GCUL nodes.

Whether GCUL is exposed to a bank’s client or used merely as a bank backend is entirely at the integrating institution’s discretion. The banks participating in the pilot chose the latter, utilizing the ledger to upgrade their backend infrastructure. With GCUL under the hood, they improved intra-day transaction performance and provided customers with the user experience they were already accustomed to.

Regulatory compliance and governance by design

To join the Swiss Franc currency network on GCUL, participants must have regulatory licenses that let them perform third-party transfers and issue commercial bank money. Crucially, features like token and account management are not rigid network-wide limitations. Instead, they rely entirely on the currency operator and its specific rulebook. This standardizes fees, limits, settlement times, and technical configurations.

Managing counterparty risk is central to this governance. Rather than relying on unscalable bilateral agreements or complex clearinghouse default waterfalls, GCUL mitigates risk through these currency operator-enforced issuance limits.

It’s important to distinguish GCUL’s broader feature set from the concrete pilot setup. In the pilot, both participating banks operated with conservative initial limits based on their credit ratings in order to safely prove the architecture. But beyond the pilot, the network frees up capital by periodically netting outstanding obligations, which frees banks from having to fund and settle every transaction.

Looking ahead, risk management can be further innovated. One clever future enhancement would be to allow participants to dynamically increase their issuance limits by committing digital assets as collateral in segregated accounts, using solutions like Crypto Finance’s AnchorNote pledging solution. These assets could then be used to instantly settle outstanding street-side obligations, virtually eliminating counterparty risk without tying up traditional liquidity reserves.

The future

While Crypto Finance’s pilot was designed to focus on domestic commercial bank money transfers in Switzerland, GCUL's programmable architecture and high-performance smart contract language, GCULpy, are paving the way for deeper innovation. As institutions like Crypto Finance scale their use of GCUL, the network can expand to support real-time cross-currency settlement, automated foreign exchange, and dynamic collateral management involving tokenized digital assets. And looking further ahead, this infrastructure has laid the groundwork for next-generation financial services that range from point-of-sale payments to autonomous, AI-driven agentic payments.

The pilot proved that deploying high-performance, blockchain-based payment rails doesn’t mean compromising on strict regulatory compliance or core banking stability. By combining the high-throughput of GCUL with a secure, API-driven integration architecture, commercial banks can offer stablecoin-like speed to their customers while seamlessly preserving their capital efficiency. Ultimately, this compliant-by-design approach targets the operational bottlenecks of traditional finance while mitigating the inherent risks and complexities of public blockchains.

To learn how your institution can build on this blueprint to fundamentally upgrade your operational payments infrastructure, contact Google Cloud Sales to explore the Google Cloud Universal Ledger.

The team would also like to thank Phil Dettwiler, Andras Miklos, Ignazio Pollina, and Ozan Kizildogan at Crypto Finance and Googlers Chris Esch, Ishan Singh, Olena Clayton, and Stathis Onasoglou for their contributions.

Expanding Claude availability and deployment choice on Google Cloud

Thu, 22 Oct 2026 21:00:00 +0000

Every organization has unique requirements for building and scaling generative AI. At Google Cloud, we are committed to giving developers model choice and architectural flexibility to run workloads however they need.

Today, we’re announcing the general availability of an Anthropic-managed deployment option for Claude Models on Google Cloud. This gives teams more choice in how they run their models, helping them consolidate AI workloads on Google Cloud alongside our existingGoogle-managed deployment option.

Both options deliver the same frontier models on day zero, the same pricing, and the same Google Cloud billing. The difference comes down to who acts as your data processor and whether you need access to the new Claude platform features when they launch:

[NEW] Claude Models on Google Cloud (Anthropic managed)

Delivers day-zero access to Claude models and Claude Platform features such as Claude Managed Agents (beta), skills, and code execution. Anthropic operates the stack as the data processor on Google Cloud infrastructure under Anthropic’s terms of service, while organizations authenticate with Google Cloud IAM, maintain Cloud Audit Logging, and consolidate spend under their existing Google Cloud billing and eligible commitments.

Learn more: Claude Models on Google Cloud (Anthropic-managed).

Claude Models on Google Cloud (Google managed)

Continues to deliver day-zero access to the latest Claude models with strict governance, compliance, and sovereignty requirements. Google Cloud operates as the data processor under existing customer agreements, providing VPC Service Controls, Customer-Managed Encryption Keys (CMEK), and data residency in the U.S., EU, and APAC. New Claude Platform features roll out over time as they are integrated into the Google-managed infrastructure. Teams also benefit from native integrations with Google Cloud services like BigQuery and Model Armor, along with standard billing and eligible commitment drawdown.

Learn more: Claude Models on Google Cloud (Google managed)

Comparing deployment options

Feature	Anthropic-managed	Google-managed
Model availability	Available on Day 0	Available on Day 0
First-party tools and betas	Available on Day 0	Available once integrated into the Google-managed infrastructure
Data processor and terms	Anthropic (under Anthropic’s Commercial Terms)	Google Cloud (under existing Google Cloud customer agreements)
Pricing & commitments	Standard pricing (draws down eligible Google Cloud commitments)

No matter how you choose to deploy, you can access Anthropic's latest frontier models on Google Cloud, including Claude Fable 5.1, Sonnet 5.5, Haiku 5.5, and Opus 5.5, with new platform features available immediately on the Anthropic-managed option and rolling out on the Google-managed option once integrated. This gives developers the full spectrum of reasoning, coding, and multi-modal capabilities with Google Cloud IAM and Marketplace billing. Get started.

Google Cloud expands model choice and flexibility, bringing OpenAI models to customers

Wed, 21 Oct 2026 12:00:00 +0000

Google Cloud has always been defined by choice, enterprise-grade governance, and market-leading price-performance. Key to the foundation of openness and choice is giving customers flexibility to run agentic workloads across their choice of models and infrastructure to achieve the right balance of cost, capability, and performance. Today, we are expanding that foundation by launching OpenAI models in Google Cloud’s model garden.

Google Cloud is now the only hyperscale cloud provider offering access to all three leading closed source LLMs, alongside more than 200 first-party, third-party, and open models on a unified enterprise AI platform.

The right model for every workload

Modern enterprise AI requires matching each workload to the right model rather than relying on a single model for every task. By adding OpenAI’s newest models to model garden, we are giving teams options across the full spectrum of enterprise workloads. For OpenAI models:

GPT-6 Astra: Designed for demanding work across coding, computer use, professional tasks, and scientific research.
GPT-6.1 Sol: Near-Astra intelligence at lower cost across agentic coding, computer use, and professional work.
GPT-6 Luna: A fast, cost-efficient model for everyday work at scale.

Using the platform, customers can dynamically select and swap the optimal model for any use case based on performance, latency, and cost efficiency without architectural lock-in or separate vendor contracts. Developers gain the freedom to build with hundreds of choices in model garden. Because OpenAI models are transacted directly through Google Cloud Marketplace, all usage appears on a single Google Cloud invoice, draws down existing cloud commitments, and counts toward Flexible Savings Plans (FSP) across model providers.

Built-in data governance and enterprise security

Scaling generative AI should never require compromising on compliance or security. Built from the ground up for enterprise governance, the Gemini platform is trusted by leading global organizations today. We ensure that running any third-party or OpenAI workload meets the most rigorous enterprise standards:

Strict data perimeters: Customer prompts, context, and outputs remain entirely within your designated data perimeter and are never used to train foundation models, with Zero Data Retention (ZDR) options available coming soon.
Isolated access boundaries: OpenAI models run on dedicated Google Cloud infrastructure with Confidential Compute coming soon. All queries are routed through hardened, isolated gateways, ensuring data transit remains strictly within Google Cloud's secure environment and third-party model providers have no unauthorized access to your cloud environment.
Unified auditing and identity: Every agent and model interaction operates under an authenticated identity, giving IT administrators end-to-end visibility, VPC Service Controls protection, and unified audit logging across their entire enterprise footprint.

“Helping customers put AI to work across their enterprise content requires leading models backed by strong security and rigorous compliance standards,” said Yashodha Bhavnani, VP of AI Products at Box. “By using OpenAI models through Google Cloud, we will be able to give customers access to OpenAI’s frontier AI models alongside Google Cloud’s security and compliance capabilities, enabling them to get more value from their enterprise content. This is a step toward operating more efficiently as we bring AI to customers.”

Getting started

OpenAI’s GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna models are rolling out now in private preview to select enterprise customers in Google Cloud’s model garden, with more coming soon. Starting today, customers can build with these models through our APIs and connect their local Codex CLI and Desktop workflows directly to our Gemini platform for inference. Please use this form to express interest.

Coming soon, OpenAI models will be available in Google Cloud’s Gemini for business and Google Antigravity, providing enterprise customers and developers with model choice flexibility and the best model for each task.

Claude Opus 5.5 and Claude Sonnet 5.5 are now available in Google Antigravity

Tue, 20 Oct 2026 16:00:00 +0000

Whether you are iterating on a quick bug fix or planning a complex migration, different tasks call for different models.

In keeping with Google Cloud's commitment to an open ecosystem and to give developers the right model for every job, Claude Opus 5.5 and Claude Sonnet 5.5 are now available in Google Antigravity for Google Cloud customers. Claude Opus 5.5 is built for complex work while Claude Sonnet 5.5 is a faster, lower-cost complement.

Once enabled by an administrator, developers can select Claude Opus 5.5 and Sonnet 5.5 directly from the model picker. Administrators can also select Google’s Gemini models, including Gemini 3.8 Flash and Gemini 4 Argon. The models apply across Antigravity 2.0, Antigravity CLI and our recently released Antigravity IDE extensions for VS Code, JetBrains IntelliJ, and Visual Studio. This lets you choose the right model for the job, under a unified enterprise billing and security boundary.

In this post, we’ll share two engineering patterns for building in Antigravity using these models. To view the entire list of models supported in Antigravity across subscriptions, see our documentation here.

Multi-model engineering patterns in Antigravity

Having multiple model families available inside the same agent harness unlocks practical multi-model workflows that improve code quality while controlling latency and cost.

Pattern 1: Cross-model verification ("LLM-as-a-Judge") in loop and graph workflows

When building agentic loops or multi-step graph workflows, using the same model to both write code and review its own output can reinforce model-specific blind spots. Pairing two distinct model families, one as the implementer and another as the validator (judge), provides independent verification before moving to the next stage of a workflow.

In Antigravity, you can run this pattern either interactively or via automated subagents:

Plan and implement with Model A: Start in planning mode using Claude Opus 5.5 or Gemini 4 Argon to design and implement a multi-file change.
Validate with Model B: Before committing or advancing the loop, switch models, or invoke a one-shot /model check in the CLI, to have a different model family critique the implementation against your test suite and requirements.

Pattern 2: Pairing reasoning models with Flash subagents

Complex engineering tasks often combine a few high-reasoning decisions with dozens of high-volume, mechanical lookups (searching files, grepping logs, checking lint output, or classifying test failures). Running every lookup on a frontier model can increase both latency and spend.

Antigravity's Subagent customization system (.agents/agents/*.md) lets you assign specific model tiers (inherit, flash, or pro) to specialized subagents. When your primary session is running a more powerful model.

Subagents configured with model: inherit use your active session model for deep reasoning subtasks.

Subagents configured with model: flash offload high-speed search, classification, and triage work to Gemini Flash models.

For example, you can define a fast log-triage subagent in .agents/agents/test-triager.md:

code_block: <ListValue: [StructValue([('code', '---\r\nname: test-triager\r\ndescription: Runs unit tests, classifies failures, and returns a concise summary of failing assertions.\r\nmodel: flash\r\ntools:\r\n - read\r\n - edit\r\n---\r\n\r\nRun the requested test targets, classify any failures by root cause (compilation error, assertion failure, or timeout), and return a structured summary to the parent agent.'), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7fc406dee950>)])]>

And a cross-model code reviewer subagent in .agents/agents/gemini-judge.md:

code_block: <ListValue: [StructValue([('code', "---\r\nname:cross-model-judge\r\ndescription: Independent LLM-as-a-judge verifier that audits diffs and edge cases using Gemini Pro.\r\nmodel: pro\r\ntools:\r\n - read\r\n---\r\n\r\nReview the current workspace diff against the user's requirements. Check for regressions, missing edge-case tests, and security issues. Output a pass/fail verdict with specific line references."), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7fc40718fbd0>)])]>

With this setup, you can delegate rapid test-and-classify loops to test-triager and trigger cross-model-judge to validate the final diff, combining the strengths of each model in a single workflow.

Getting started

Claude Opus 5.5 and Claude Sonnet 5.5 are available now for Google Cloud customers using Antigravity:

Administrators (Gemini for business): Third-party partner models do not draw from the pooled Gemini Enterprise Standard and Plus edition monthly token quota. Instead, usage is billed on a pay-as-you-go basis through your existing Google Cloud Billing account, so overages must be enabled. Visit “Gemini Enterprise” > “Settings” > “AI developer tools” in the Google Cloud Console to configure overage billing, spend limits and authorize partner models for your organization. See Configure AI developer tools settings for step-by-step instructions.
Administrators (Gemini platform): You can also enable partner models if you are using Antigravity through the API in Gemini platform. Enable Claude models in Model Garden and complete the standard Anthropic terms.
Developers: Update to the latest version of Antigravity 2.0, the Antigravity CLI, or the Antigravity IDE extensions, and open the model picker (or run /model) to select the right model for your task. To learn more about enterprise governance and multi-agent configuration, visit the Google Antigravity Enterprise documentation. You can switch the active model between turns in the same conversation thread while preserving workspace state, open artifacts, and terminal context.

Note: Lightweight first-party Gemini models continue to run in the background to handle high-frequency client-side coordination, such as context management and summaries. They consume first-party tokens and can incur charges.

Introducing Gemini for government services

Tue, 20 Oct 2026 10:00:00 +0000

Earlier this month at Gemini at Work, we introduced the Gemini agent: a single, universal agent for work. It answers your questions, handles your knowledge work, creates your images and media, and writes and runs code. You give it objectives, not instructions, and come back to finished work. It plans the work, uses skills and tools, connects to your systems, and has the security, administration, and governance your organization requires.

Every industry runs on knowledge a general model does not have. Government work runs on statute, regulation, and agency guidance, across siloed legacy systems, under strict requirements for data sovereignty, compliance, and confidentiality. Too often that leaves civil servants on data entry and manual document review instead of the services citizens depend on.

Today we're bringing the Gemini agent to public service with Gemini for government services in preview, part of our new suite of purpose-built industry solutions. It adds skills built around government work, connections to the systems agencies already run, and specialized partner agents, with governance running underneath it all.

Five components of Gemini for government services

Developed alongside agency leaders, Gemini for government services provides an integrated, fully governed environment configured for rapid deployment across federal, state, and local agencies with five core components:

Skills for government work: Skills are reusable instructions and context, written by domain experts, that teach Gemini to run a specific task. Each one applies the statute, regulation, and agency guidance that govern that task, so an agency's standard operating procedures carry into the work Gemini supports.
Connectors to agency systems: Secure connectors link Gemini to the case management, ERP, HR, and collaboration systems agencies already run, along with authoritative federal data sources. Gemini queries each system in place, so no data is migrated, and access follows the identity and permission controls the agency already has.
Partner agents: Skills and connectors come together in agents, including pre-built agents from public sector software providers that agencies can deploy right away. They handle tasks like FOIA discovery and batch redaction under the same central governance, with a person approving each consequential step.
A governed control plane: Security and compliance teams get one place to set and enforce policy across every agent, including VPC Service Controls, CMEK, and Assured Workloads. Outputs are grounded in source documents, with citations, so reviewers can check where each answer came from.
An open partner ecosystem: Every agency works under a different statute and runs a different mix of systems. System integrators and public sector ISVs help agencies customize Gemini, connect it to what they already have, and scale it without lock-in.

Skills for government workflows

Gemini for government services helps automate and accelerate complex government operations, including:

Policy and regulatory research: Ask an operational or eligibility question and get an answer drawn from the CFR, the Federal Register, state administrative codes, and internal manuals, with page and paragraph citations. Where federal and state rules both apply, Gemini shows which one governs to empower agency decision makers.
FOIA record discovery: Gemini searches SharePoint, Drive, and email archives for records responsive to a FOIA or state Sunshine Law request. It organizes them into a review set for the records officer and agency counsel, shortening discovery that often takes weeks.
Redaction for public release: Gemini finds sensitive constituent data across a batch of documents, groups the findings so reviewers can clear them in passes, and proposes redactions labeled by statutory exemption. After a reviewer confirms, it produces the exemption log and audit-ready copies.
Vendor bid evaluation: Gemini checks multi-volume bids against RFP requirements, FAR guidelines, minimum mandatory criteria, and evaluation rubrics, and returns scoring justifications with citations. Contracting officers get a documented basis for each award decision.
Grant and program expense audits: Gemini checks invoices, receipts, purchase orders, and ledger entries against 2 CFR 200 Uniform Guidance and the terms of the award. It flags unallowable costs, missing documentation, and funding discrepancies with line-item citations, so problems surface while there's still time to fix them.
Document validation at intake: Gemini classifies and validates pay stubs, W-2s, utility bills, and leases when they arrive, checks the numbers for consistency, and flags missing signatures. Defects are caught before the application enters the queue, not weeks into the case.
Plain-language notices: Gemini turns policy memos and agency guidance into public notices, aligned to the Plain Writing Act of 2010, USDA-FNS standards, and Title VI language access guidance.

Open ecosystem of connectors across the government technology stack

Gemini for government services connects directly to core agency systems via secure connectors. Access is bound by existing role-based access controls, record-level permissions, and the trusted data controls inherited from each system — and nothing is migrated to make the connection work.

Federal regulatory and public data:

Federal Register: Tracks legislative actions, regulations, and executive orders as they post, so state and local teams see federal changes as they land.
regulations.gov: Tracks federal rulemaking dockets, public comments, and statutory agency response deadlines, keeping policy analysts ahead of incoming mandates.
eCFR (Electronic Code of Federal Regulations): Grounds answers in the governing regulation, with citations down to the paragraph.

Agency systems of record:

Oracle E-Business Suite: Gives agents permission-bound access to procurement, payables, and ledger transactions for expense auditing and award reconciliation.
Oracle PeopleSoft: Connects HR and personnel records under existing role-based controls, so workforce questions are answered from the system of record.
Salesforce: Enables conversational experiences across the Missionforce portfolio through headless integration with Gemini, including benefits, licensing, permitting, and contact center management.
ServiceNow: Automates IT, CRM, security, employee experience, and workflow records, so agents can track incidents, requests, and approval chains inside existing routing rules.

Productivity and collaboration:

Google Workspace: Connects with Google Docs, Gmail, Drive, and Sheets to analyze case communications and agency correspondence and surface internal files while enforcing agency access controls.
Microsoft 365 and SharePoint: Integrates with Word, Outlook, and SharePoint to triage inquiries and ground responses in mail and intranet repositories without breaking workflow context.

Third-party agents and implementation partners

Every agency operates differently. Through our open platform, agencies can deploy pre-built partner agents or work with system integrators to scale custom capabilities:

SZNS Solutions FOIA and public records discovery agent: Turns a plain-language request, a piece of correspondence, or a formal citizen inquiry into a scoped search across enterprise archives — bounded by custodian and date range, so a records officer narrows the set before any data moves. It exports responsive material into secure agency repositories and tracks delivery status, giving each request an auditable trail from intake to release.
SZNS Solutions document redaction agent: Works inside connected repositories — SharePoint and Drive among them — finding sensitive citizen data across a batch and grouping the findings so a reviewer clears them in passes rather than page by page. Each proposed redaction carries its statutory exemption, and once the reviewer confirms, the agent exports permanently redacted, audit-ready copies back to the system of record, with the processing never leaving the agency's cloud perimeter.
Systems integrators & tech partners: Strategic partnerships with Accenture Federal Services, Clutch, Dark Wolf, Daston, Deloitte, Egen, Game Plan Tech, Knexus, Onix, Resultant, SAIC, Slalom, and SZNS Solutions so agencies can customize, integrate, and scale these capabilities without vendor lock-in.

Developed alongside federal, state, and local government agencies

We are working closely with select federal and state agencies, including the State of Missouri and the U.S. Department of Health and Human Services to ensure these capabilities work in day-to-day agency operations.

"Our citizens expect the government to move at the speed of modern life, not the speed of paperwork. By automating routine work in our workflows, we're delivering faster, better services to Missourians." - Tim Marczewski, Director of AI and Innovation, State of Missouri Office of Administration.

Built on an enterprise-grade foundation

Gemini for government services runs on Google Cloud's secure foundation, and can be deployed using the Assured Workloads FedRAMP High Data Boundary, enforcing U.S. data residency and personnel controls. It works within the authorities, retention rules, and approval chains your agency already has.

Agency records, constituent data, institutional documents, custom agents, and model outputs stay private to your organization, and are never used to train or fine-tune Google's foundation models without your permission. Because we operate the full stack, from infrastructure and models through the application layer, we can tune performance and cost together — routing high-volume document parsing and routine records sweeps to fast, low-cost models and reserving high-reasoning models for complex policy analysis, so expanding what your teams can take on does not mean expanding spend at the same rate.

This is just the beginning

The launch of Gemini for government services represents another defining step in delivering on our promise to bring the best of Gemini to every professional, for every workflow, natively tailored to the way they work.

Gemini for government services is available in preview today, joining our purpose-built solutions for industries. We invite leaders at every level of government to explore how Gemini can transform their most critical workflows.

Who judges the judge? Three helpful ways to audit your LLM evaluators

Mon, 19 Oct 2026 16:00:00 +0000

Is your LLM-as-a-judge truly objective?

There’s a common misconception in tech that humans are subjective, so we should rely on AI for objective evaluation. But the reality is that LLMs are also subjective. Just like human reviewers, they can carry biases:

Verbosity bias: LLM judges usually favor longer, confident-sounding answers, even when the underlying facts are wrong.
Position bias: Swapping the order of two options can cause an LLM judge to flip its verdict over 75% of the time.
The halo effect: Because LLMs generate text piece by piece, clean grammar and formatting can trick an AI judge into giving top marks for accuracy while missing errors.
Agreeableness bias: LLM judges are great at recognizing valid work. But with detecting errors, they may fail to catch incorrect, buggy, or unsafe outputs up to 75% of the time.
Self-enhancement bias: Also known as egocentric bias, AI evaluators systematically overrate outputs that were generated by themselves or from their own family.

Ultimately, LLM judges are not objective arbiters of truth. They are subjective models with their own distinct biases. Treating them as infallible tools that don't need their quality measured might leave your work compromised.

In this post, we’ll explore what’s holding enterprises back from strong evals, and share an easy three-pillar method you can use to improve your quality.

Three helpful ways to audit your LLM judges

Pillar 1: Measure internal stability

Before you can trust an AI judge’s score, you need to know that it isn't just guessing. Because LLMs are probabilistic, they can be inconsistent, giving different scores to the exact same text.

You can test this internal stability by running multi-pass consistency checks on a single output and calculating the variance. If a judge's scores fluctuate wildly on identical inputs, its rubric is too ambiguous.

Pillar 2: Validate human alignment

When your evaluation datasets are mostly valid responses, a lenient AI judge can look great on paper. It might align closely with human reviewers overall, but miss a majority of invalid or even unsafe outputs.

To prove your LLM judge is catching real errors, you need chance-corrected metrics. These mathematical models filter out agreements driven purely by leniency or random chance, showing how closely your evaluator actually aligns with the informed decisions of human experts.

Pillar 3: Use multi-judge ensembles

To avoid the halo effect mentioned earlier, teams should use decomposed prompting. Instead of asking one LLM judge to score an entire response, assign a specialized judge to evaluate a single criteria, like factual accuracy or tone. This isolates traits and forces independent scoring.

Second, to track performance in real time without human intervention, the established best practice is ensembling. This means running at least two different models under the hood to evaluate the exact same output simultaneously.

By finding the statistical consensus between these models, you can flag if their evaluations begin to diverge and avoid any scoring quality degradation.

Get started

Partner with Delta to design and implement statistically-audited evaluation pipelines for your enterprise agentic systems.

Building full virtual vehicles on Google Cloud: From AAOS to the smallest MCU

Mon, 19 Oct 2026 16:00:00 +0000

Imagining traditional gearheads getting their hands on software-defined vehicles is an amusing thought experiment. With their socket wrenches and air compressors at the ready, they might be confused to encounter core functions more likely to run on 1s and 0s than oil and belts. But across the auto industry these days, more and more vehicles feature prominent systems built using cloud-native development, continuous integration/continuous delivery, fast digital iteration, and plenty of AI.

But there’s still a big gap to be found in many automakers’ and Tier 1 suppliers’ cloud-based development processes. On the one hand, these companies are commonly using software to develop vehicles’ digital entertainment and information screens, or other onboard high-performance computers.

On the other, the dozens of microchips and control modules that make a car a car these days — such as gateways, body controllers, zone control units, sensor and actuator nodes — are still frequently hardwired to physical chips when they’re constructed, making them very difficult to update, reprogram, or expand after the vehicle drives off the factory floor. Teams have to wait for benches, samples, and rigs, and integration can easily slip to the end of the development process, where defects are most expensive to fix.

This inherent inefficiency has become a leading bottleneck. After all, manufacturers can’t really claim vehicles are software-defined if most of their electronics only function when connected to physical hardware and if developers are leaning heavily on actual hardware.

With Horizon, our agentic innovation platform for developing software-defined vehicles, our vision is to help manufacturers avoid these inefficiencies by working with partners like RemotiveLabs.

The cockpit already runs in the cloud

The Android team has invested a significant amount of time and effort over the last few years on enabling seamless development of the many digital systems in vehicles’ cockpits. Android Automotive OS (AAOS) runs inside Google’s official virtual testing environment for Android, Cuttlefish, which itself runs comfortably on Arm-based Google Cloud instances. With this approach, developers can build, boot, and test virtual versions of their infotainment stacks on the same Arm architecture that ships in vehicles.

And with the recent launch of Android 17 and AAOS SDV, this capability is also extended to further domains, whenever AAOS SDV is used. Deploying it with Cuttlefish, for example, proves the model: representative software, on representative architecture, in the cloud, on demand. The question that still needs to be answered is how to extend that model to the rest of the vehicle, especially when other operating systems are involved.

The harness around the virtual vehicle

That’s where RemotiveTopology comes into play. As the harness around the virtual vehicle, it’s the layer that connects AAOS in Cuttlefish to a full, simulated electrical and electronic (E/E) architecture: the rest of the vehicle’s onboard computers, its data networks, and the live signals that make the system function like a real vehicle rather than a collection of standalone apps.

RemotiveTopology automatically utilizes manufacturers’ existing network design files to build a working, virtual version of their cars’ networks. This is vehicle-as-code: the E/E architecture expressed as an executable specification, versioned and provisioned like infrastructure-as-code. In this way, an entire vehicle can be spun up on demand. Cockpit software talks to the rest of the vehicle exactly as it would in a real vehicle. As a result, developers get full and continuous signal and cross-domain visibility, so they can see real-time system status. And because it is built on standard Linux tooling, the same topology runs on a laptop and in a contiguous integration pipeline on Google Cloud.

Integrate across domains early, not late

The payoff for using this approach is that all onboard systems can be developed and integrated with software before any of them are connected to a vehicle’s hardware. That’s beneficial because the most difficult (and latest) integration in a vehicle’s development is bringing all its systems together. For example, advanced driver assistance systems (ADAS) produce object lists, trajectories, and mode requests that the cockpit has to render and react to, while the infotainment system and information cluster have to reflect driving state while body and chassis nodes sit in the loop.

With the full topology live on Google Cloud, the infotainment system and ADAS can operate simultaneously within a virtual vehicle, exchanging real signals over the simulated network. This means it’s possible to exercise handoffs like an ADAS event surfacing in the cluster or a driver-mode change rippling across domains from day one in continuous integration long before the benches and samples line up. In this way, developers can be confident they’ll discover multi-domain integration long before the end of the process so that they can test continuously from the start.

The full vehicle frontier

At the heart of this work is a shared desire to bring scalable software development to even the smallest and most common nodes in a vehicle, its microchips and control modules. That’s because these are the parts that have traditionally been wired to hardware, and as a result are the biggest blocker to fully software-defined vehicles.

The ultimate goal is that the same software that runs in the cloud can run on vehicles’ actual control modules with no re-tooling or re-validation needed to move between them. That parity is natural with Arm-based Google Cloud infrastructure because the cloud and the car share an instruction set. Google's Axion makes it straightforward: C4A metal, a bare-metal Arm instance, is provided explicitly for automotive systems development and Android software development, giving direct hardware access without unnecessary software layers and delivering the consistent performance that vehicle systems require.

Taken together, this holistic system allows teams, for the first time, to develop and integration-test an entire vehicle’s systems in the cloud and then apply it to the actual vehicle without having to rewrite it.

Why now, and where to start

The payoff is not abstract. Treating the vehicle as code and moving integration into the cloud compresses the slowest, most expensive phase of a program. Early customer engagements with RemotiveLabs have measured more than 10x faster integration cycles. And the development loop is now changing faster than the hardware loop: AI-assisted and agentic workflows can rapidly generate, test, and refine vehicle software, but only virtual systems provide the required scale. Every element that can’t be virtualized can slow the entire development process down.

RemotiveTopology provides this architecture as vehicle-as-code and makes it accessible for developers, testers, and autonomous AI agents alike. Best of all, there is no new third-party tooling to adopt just to get started. Instead, RemotiveLabs publishes an open-source simplified vehicle on GitHub that anyone can run in minutes on standard Linux tooling and platform definitions already in place. Full existing toolchains plug in as a manufacturer or Tier1 supplier scales, and RemotiveTopology can be easily dropped into an Arm-based Google Cloud environment.

For those building software-defined vehicles whose virtual development stops at the cockpit, it’s time to start with an open-source example, bring unique platform definitions, and run an entire vehicle in the cloud. RemotiveTopology is available in Horizon for easy usage in combination with other Horizon modules.

Google named a Leader in the 2026 IDC MarketScape for Agentic AI Platforms

Mon, 19 Oct 2026 07:00:00 +0000

As organizations work to leverage AI to transform their businesses, it is clear that the raw model intelligence is not the key challenge.

Consistently, we hear that the challenges lie in how the agents can be deployed across the company while meeting key requirements for quality, integration, security, and governance.

Gemini provides that connected foundation across the enterprise, and it’s why we believe Google was named a Leader in the inaugural IDC MarketScape: Worldwide Agentic AI Platforms 2026 Vendor Assessment¹.

IDC MarketScape vendor analysis model is designed to provide an overview of the competitive fitness of technology and suppliers in a given market. The research methodology utilizes a rigorous scoring methodology based on both qualitative and quantitative criteria that results in a single graphical illustration of each supplier’s position within a given market. The Capabilities score measures supplier product, go-to-market and business execution in the short-term. The Strategy score measures alignment of supplier strategies with customer requirements in a 3-5 year timeframe. Supplier market share is represented by the size of the icons.

Gemini meets work where it already happens

A viable agentic platform has to meet work where it already happens – and run under production-grade controls. Gemini is engineered to deliver developer velocity without compromising governance, built around four core differentiators:

1. Surfaces that connect builders and business teams

Gemini makes it so that every surface connects back to one universal foundation.

Where teams already work: Employees collaborate with Gemini directly inside Google Workspace, Microsoft 365, or team chats – the agent carries the same persistent memory, skills, and enterprise permissions across desktop and mobile without needing to switch tools.
Builder tools centered on choice: Gemini platform supports how developers want to work, spanning no-code and low-code builders in the studio to code-first workflows with the Agent Development Kit (ADK) and Antigravity.
Workflows that compound across the organization: When a team solves a complex workflow, they can package it into a shared skill published to a central Agent Registry. Gemini calls on these skills directly to execute tasks, turning proven procedures into reusable assets across the entire company.

2. Radical openness across models, enterprise data, and tools

Gemini adapts to the software you already run, not the other way around.

Model choice: Through Model Garden in Gemini platform, teams have direct access to over 200 models – including all of the breakthroughs from Google DeepMind like Gemini 4 Argon, day one access to the latest Anthropic models, and third-party and open models.
Grounding on your enterprise data: Gemini connects directly to live enterprise data across Google Cloud in BigQuery and Cloud Storage, as well as Salesforce, SAP, and others. Because Gemini inherits each employee's existing credentials, sensitive information stays governed at the source.
Open standards: Built on open protocols like MCP and A2A, Gemini can easily plug into your existing toolchains and communicate across multi-vendor environments.

3. Smarter token economics

Gemini optimizes cost, latency, and performance directly at the execution layer.

Matching tasks to the right model: By default, Gemini uses Auto mode for users, letting the platform match each step of a task to the right model dynamically. It defaults to a fast workhorse model like Gemini Flash, and escalates to a high-capacity frontier reasoning model only when task complexity requires it.
Reusable skills: Teams can package proven processes into shared skills. Your agents can execute against these established procedures, improving consistency and reducing repetitive token consumption across the company.
Budget guardrails: Administrators can set spending limits by project, department, or individual agent in the billing console. When a budget threshold is reached, new requests pause while active tasks are allowed to complete, giving finance and engineering predictable cost controls without abruptly interrupting in-progress work.

4. Built-in governance and security

Gemini incorporates enterprise governance natively, and at no additional cost.

Dedicated agent identity: Every agent receives its own distinct, authenticated identity governed by least-privilege permissions. When an agent takes action, every event is recorded in an audit trail tied directly to that identity and the identity of which user was using the agent for the task.
Isolated sandboxes: A sandboxed boundary keeps dynamic tasks strictly contained, protecting your underlying infrastructure, core network, and sensitive corporate data from being exposed or compromised.
Policy enforcement: Rather than relying on a model to self-police, Agent Gateway automatically applies corporate security policies, blocks data leakage, and ensures agents only access what they are explicitly authorized to touch.

Looking forward

At its core, Gemini is all about giving your organization an open, reliable platform where builders have choice and teams work with complete confidence.

We are honored by IDC MarketScape’s recognition of Google as a Leader, and we look forward to continuing this work alongside our customers.

Download the IDC MarketScape: Worldwide Agentic AI Platforms 2026 Vendor Assessment excerpt to learn why Google is named a Leader.

^{1. SOURCE: “Worldwide Agentic AI Platforms 2026 Vendor Assessment”, October 2026, IDC # US54140926.}

Architecting autonomy: A design pattern for high-compute, governed agents in biotech

Fri, 16 Oct 2026 16:00:00 +0000

In production systems today, most AI agents operate as lightweight dispatchers. They parse a prompt, make an API call, retrieve a record, and hand back control. Their execution is fast, stateless and cheap. Biotechnology does not work this way.

Folding a protein target, screening a 50,000 compound ligand library, or evaluating structure-activity relationships (SAR) are not quick API lookups. They are sustained, stateful, GPU-bound simulations that run for minutes or even hours across specialized silicon. When you ask a general-purpose large language model (LLM) endpoint to coordinate a complex drug discovery workflow, the standard paradigm breaks: Timeouts occur, context windows overflow with atomic coordinates, and proprietary chemical Intellectual Property (IP) risks exposure to external networks.

Building enterprise grade research and development (R&D) agents requires moving past the monolithic assistant model. In this blog post, we will discuss how to use a three-layer architecture: separating cognitive orchestration (the brain) from computational execution (the muscle), and anchoring both inside an enterprise governance plane to build scalable agents for drug discovery use cases in biotech.

A three-layer pattern

The reference architecture below shows how to architect, optimize, and govern high-compute autonomous agents in biotechnology, specifically Drug Discovery.

Figure 1 — The three-layer pattern: a cognitive orchestrator (brain), a scalable GPU execution layer (muscle), and an enterprise governance plane.

The execution layer: scaling the muscle - high-compute execution

The execution layer is responsible for the heavy scientific computations. The first architectural decision is determining where workloads should run:

Strategic, proprietary pipelines require dedicated GPU infrastructure within your security perimeter
Routine or standardized workloads can safely be delegated to managed cloud endpoints

In this architecture, protein folding is driven by ESMFold and generative design by GenMol. Both run as NVIDIA Inference Microservices or NIMs deployed on Google Kubernetes Engine (GKE) using G4 VMs backed by NVIDIA RTX PRO 6000 Blackwell GPUs.

A critical hurdle in containerized GPU workloads is the cold-start penalty, the latency delay required to pull multi-gigabyte model weights from an image registry and stage them into GPU memory (VRAM).

To ensure low latency responsiveness, we build the following as a recommendation into the reference architecture:

Maintain warm pools: Pin the GKE node pool to a minimum of one active node.
Preload weights: Model weights are loaded directly into VRAM at pod initialization, ensuring containers are immediately ready for inference.
Resilient scheduling: GKE manages pod lifecycles and restarts failed instances without tearing down the underlying VM infrastructure.

This approach trades custom, manual CUDA kernel tuning for NVIDIA's pre-compiled, containerized runtimes, dramatically reducing infrastructure engineering overhead while ensuring deterministic throughput.

In biotech, chemical structures represent millions of dollars in proprietary IP. A docking request transmits SMILE strings or 3D coordinate matrices that must never leave company custody.

Containerized NIMs eliminate the friction between rapid prototyping and rigorous data isolation. During early exploration, researchers can query NVIDIA hosted endpoints. When transitioning to proprietary lead candidates, the exact same container is deployed to GKE inside a private VPC. The agent’s orchestration logic remains identical; only the target URI updates from an external proxy to an internal VPC endpoint. No structural data leaves the enterprise perimeter.

The cognitive layer: Multi-agent reasoning

With compute execution isolated, the cognitive layer manages agentic workflow sequencing. Built on Agent Runtime, within the Gemini platform, this reference architecture shows a system that deploys a hierarchical multi-agent structure, one lead orchestrator, and two specialized domain sub-agents (Lead Optimization and Safety Screening) . The multi-agent design can be modified to suit the use case at hand.

Based on this design, when an end-user submits a high-level task such as “ Analyze Candidate 4b against Target X and produce a SAR report”, the system applies four core orchestration patterns discussed below;

In-process native agent handoffs: The lead orchestrator delegates tasks to specialist sub-agents using Agent Runtime’s native SDK rather than network-serialized agent-to-agent HTTP protocols. This eliminates intermediate network hops and keeps internal handoffs instantaneous. Only the outermost boundary communicates outward to user-facing surfaces like the Gemini agent.
Passing bulky payloads by reference: LLMs degrade quickly when forced to ingest massive arrays of 3D spatial coordinates. In this architecture, raw structural data (PDB files, docking poses, trajectory logs) is written directly to Cloud Storage. Agents pass lightweight, signed object URIs and high-level structural summaries, keeping context windows compact and inference costs down.
Decoupling workflow logic from biochemical reasoning: General-purpose LLMs excel at task planning and synthesis, but are prone to hallucinating biochemical nuances. To control this behavior, this architecture maintains strong separation of concerns, namely, workflow and tasks decomposition are handled by Gemini. Domain chemistry and pharmacological reasoning are handled by specialized Nemotron models hosted on dedicated GPU nodes.
Context-driven evaluation over rigid cutoffs: Traditional computational screening enforces arbitrary numerical cutoffs (e.g., discard if binding affinity ,-7.5 kcal/mol) However, therapeutic context fundamentally alters risk tolerance. Preventive cardiovascular therapy taken daily by healthy patients requires extremely strict safety margins, or a novel oncology candidate targeting an intractable cancer may accept higher toxicity if efficacy is demonstrated.

By passing disease context directly into the specialized reasoning models, drug candidates are evaluated against dynamic, target-specific profiles rather than static thresholds, preventing the premature culling of viable drug leads.

Performance best practices for high-compute pipelines

Optimizing an agentic system that interfaces with heavy GPU compute requires challenging common engineering instincts:

Common Assumption	Architectural Reality	Recommended Best Practice
"Swap in a faster driver model to cut latency."	The primary orchestrator's token generation is rarely the bottleneck in multi-step R&D pipelines.	Optimize I/O, not token speed: Focus on reducing payload sizes and intermediate network hops rather than compromising orchestrator reasoning depth.
"Parallelize tool calls across workers."	External scientific services and containerized models often serialize requests internally on dedicated hardware.	Respect execution topology: Align client-side concurrency with backend queue mechanics to prevent pipeline lockouts and connection timeouts.
"Pass full simulation states between tools."	Large JSON bodies cause schema validation failures, prompt truncation, and token bloat.	Pass heavy payloads by reference: Store raw coordinates in object storage and pass lightweight URIs between agent boundaries.
"Return outputs only when the pipeline finishes."	Long multi-minute runs leave scientists blind to whether jobs are proceeding or stalled.	Stream intermediate stage completions: Emit progress updates at each completed milestone (e.g., folding complete, docking running) for real-time visibility.
"Chain multiple granular micro-tools."	Deep tool-chains increase the probability of hallucinated arguments and schema mismatches.	Collapse compute pipelines: Wrap end-to-end simulation pipelines behind a single, clean tool interface that takes clear parameters and returns a consolidated object.

The enterprise governance and asymmetric autonomy

Deploying high-compute agents in enterprise environments requires robust data sovereignty, audit trails, and process control. The Gemini platform establishes these safeguards natively with the following components:

Agent Identity: Every autonomous agent operates with a cryptographically verifiable identity. Thus, for a biotech use case, when a candidate molecule is advanced or rejected, the event is permanently logged to that specific agent principal.
Agent Gateway: This serves the single ingress/egress point for every interaction within the perimeter. Centralized IAM and regional residency rules help ensure sensitive molecular data remains within designated cloud regions.
Agent Registry: This feature maintains a curated catalog of approved agents, validated tool schemas, and authorized Model Context Protocol (MCP) servers, preventing rogue tools from executing in production.
Model Armor: This functionality inspects incoming prompts and outgoing payloads in real time to prevent prompt injection and data exfiltration.

The payoff is clear. Identity, RBAC, auditing, residency, and data protection are native properties of the Gemini platform; this allows enterprise customers to focus on science rather than having to build a compliance stack for every project.

In high-stakes research and development, full autonomy introduces unacceptable liability, while requiring approval for trivial steps can create frictions. The solution addressing this to use asymmetric gating in the following ways;

Automate reversible operations: For example, allow the agent to resolve chemical naming synonyms, pull structural PDBs, and run preliminary Lipinski screens autonomously without interrupting the researcher.

Gate irreversible operations: For instance, initiating multi-GPU docking run, advancing a lead compound to wet-lab synthesis, or archiving a therapeutic program must require explicit human sign-off.

The pattern, and the path forward

High-compute autonomous agents succeed when the boundaries between cognitive planning, numerical computation, and system governance are clean in the following ways:

Keep the orchestrator focused on decomposition and synthesis, leaving intensive biophysical simulation to specialized, containerized microservices and models.
Eliminate context bloat by passing complex chemical data by reference through cloud storage.
Anchor the workflow in enterprise governance, pairing cryptographic agent identity with asymmetric human oversight.

As generative and structural biology models expand toward whole-cell modeling and complex dynamic simulations, this decoupled architecture ensures research systems scale efficiently while keeping proprietary IP secure and auditable.

Reach out to your Google sales representative if you have any requirements to build a use case like this and we will be happy to talk and support you in your agent development journey on Google Cloud.

Four evaluation engineering lessons learned from building agent plugins

Fri, 16 Oct 2026 16:00:00 +0000

The real challenge of building an agentic system isn't the construction, it's the measurement.

Evaluation engineering is a critical foundation for building reliable AI systems. Setting up systematic, repeatable tests to measure your agent’s true behavior can be the difference between shipping a tool you can trust and one that silently breaks, fails to call APIs, or makes mistakes.

I’ve run evaluation engineering initiatives for years at Google and developing the newly released Google Cloud Developer plugin served as a live testbed to stress-test these key principles across real-world developer workflows.

Here are four eval lessons we learned from building the new plugin that you can apply to your own builds.

1. Grade the execution steps instead of the summary

When we first evaluated agents, we set up an LLM-as-judge to inspect the agent’s final text reply and scored it against a checklist. However, we quickly realized it’s fundamentally flawed because the grader only sees what the agent wrote, not what it actually did. When we ran our initial 11-prompt evaluation suite for the Google Cloud Developer plugin, the automated grader marked it as failing three prompts. But when we opened the raw tool-call logs and read the transcripts, the plugin had actually passed all three. On a Cloud Run cleanup prompt, for instance, our rubric required checking command syntax before running anything. The agent ran gcloud help run services list as its very first step. It did the work, but its final message was focused on warning the user about permanent data loss rather than reciting that it had checked the help docs.

The grader penalized the agent for doing the right thing simply because it didn’t brag about it in the summary. If your rubric line says "did it run X?", a LLM judge is guessing unless you evaluate the actual tool-call trajectory.

Takeaway: Don't grade an agent on its closing pitch. Score the raw execution trajectory to verify what commands were actually executed, validated, and scoped.

2. Design evals for asynchronous operations

When you start reading tool-call trajectories, how you extract that telemetry matters just as much as what you're reading. Extracting agent telemetry into a flattened, sequential list of events obscures concurrent tool dispatch, which is one of the critical causes of behavioral failure. Modern agent harnesses often batch multiple tool invocations into a single assistant turn using an array of calls.

In our evaluations, when an agent was asked to inspect an environment and look up documentation, the harness dispatched the skill-loading call and two documentation searches simultaneously within a single tool_calls payload. Because activating a skill is a multi-stage process, where the initial call returns a short launch stub and the subsequent turn injects the full instruction body, the instructions meant to govern the search arrived in the prompt context after the system had already executed the searches.

The system never gave the skill a dedicated turn to let its guidance influence the agent's actions. If telemetry pipelines strip out turn structures and merge calls into a flat timeline, engineers will waste days wondering why an agent disregarded unambiguous instructions that a developer embedded in a skill.

Takeaway: Preserve turn structure in telemetry extractors so you can identify when client-side parallel dispatch issues tool calls before injected guidance enters the context window.

3. Test live runs and "dry runs" side-by-side

You need to inspect real API calls to catch harness quirks, like parallel dispatch. But if you only test your agent in one environment condition, you're flying half-blind. From day one, you need to run two conditions side-by-side: a live run and a plan-only run. In the live condition, the agent has real, sandboxed credentials and actually executes commands. This is how you see if syntax works, parameters serialize cleanly, and the agent exercises restraint before doing something destructive.

Live runs frequently hit an unexpected permission error on step two and abort before you ever reach the behavior you’re trying to test. That's why you need the plan-only condition, where you strip credentials and tell the agent to explain what it would do without running anything. This strips out permission roadblocks and shows you the model’s pure reasoning and routing intent. Roughly half our most important findings came from the delta between the two.

For example, on one of our onboarding prompts, a live run was blocked immediately on a project list call. The plan-only run sailed right through and produced an 8,000-character plan that completely ignored organization policies, admin approvals, and governance. The live run masked a major cognitive flaw that the plan-only run exposed in seconds.

Takeaway: Run live and plan-only evaluations side by side. Live runs test actual execution and safety restraint; plan-only runs expose the model's raw reasoning and routing without getting tripped up by permission errors.

4. Use “golden sets” as durable baselines

Frontier model architectures change every few months, agent harnesses improve, and plugin manifest standards will inevitably evolve. When building the Google Cloud Developer plugin, we treated evaluations as continuous integration testing for these non-deterministic environments. That means we stress-tested the composite stack, including the harness, the live model, and the tools running together on realistic tasks.

Instead of guessing from the final text output, this approach captures the whole trajectory: what the agent thought, what commands it actually ran, and whether it solved the problem without breaking anything. Then you score that run against golden sets curated by real human experts who know what the ideal path looks like.

Takeaway: Build golden sets based on what the entire system should actually do in the wild, including your harness, model, and plugins. Models change, but those baseline runs are your safety net.

Reminder: Use evaluation suites over “one-shot” demos

A common mistake I see among engineers is what I like to call the one-shot demo. This is when someone feeds a prompt into a LLM, watches it build a complete 3D game or solves a Rubik’s cube in a single pass and calls it an eval. But in reality, this is just a vanity metric.

To do quality work with AI, I truly believe that you need evaluation suites. Without evaluations, teams are like literally running into a dark cave blind. Engineering an enterprise-grade agent means testing the system's ability to handle multi-step interactions, manage permissions, and avoid destructive outcomes like accidentally deleting a project because you asked it to "find opportunities to save costs."

Where to go from here

Building an agent plugin or MCP server is the beginning. Shipping something you can actually trust with real cloud infrastructure requires significantly more work in evaluations to ensure it’s doing what you expect it to do. At the end of the day, the only thing that outlasts model churn is a stress-tested eval suite. Check out the Google Cloud Developer plugin documentation or explore this codelab to get started with the plugin in Antigravity.

AI & Machine Learning

The Blueprint: Orchestrating 5,000+ AI assets for a World Cup TV spot

The Challenge

The Solution

The Architecture

The Outcome

The Details

Bringing conversational analytics to Gemini with Agentic Data Cloud connectors

A new way to query enterprise data

How Data Cloud connectors work

Choosing between connectors and data agents

Built-in access controls and governance

How to create Data Cloud connectors

Why use Data Cloud connectors over custom MCP

Get started today

Gemini Omni 1.2 Flash: Fast, cost-effective, and reliable video generation and editing now generally available

Gemini Omni 1.2 Flash is ready for production

Video creation inside the tools people already use

Product and listing videos at catalog scale

Create in Gemini for business, Google Vids and build with the API

Get started with Gemini Omni

How Nammu21 automates syndicated loans on Google Cloud Universal Ledger

Digitizing the loan market

A purpose-built stack for digital credit

Technical architecture: The loan lifecycle on GCUL

The business impact of moving on-ledger

Introducing Physical-Engineering’s Last Exam for agents

Why textbook benchmarks fall short for physical engineering

Inside PLE: Real-world tasks built with practicing engineers

All frontier models score below 50%

Test your agent and join the open competition

What Google Cloud announced in AI this month

Top announcements

Thought leadership (editor’s pick)

News you can use

September

Top announcements

Thought leadership (editor’s pick)

News you can use

August

Top announcements

Thought leadership (editor’s pick):

News you can use:

July

Top announcements

Thought leadership (editor’s pick):

News you can use:

June

Top announcements

Thought leadership (editor’s pick):

News you can use:

May

Top announcements

Thought leadership (editor’s pick):

News you can use:

April

Top announcements

News you can use

March

Top hits:

News you can use:

February

Top hits

News you can use

Janurary

Top hits

News you can use

What Google Cloud announced in AI this month - 2025

What’s new in AI infrastructure and orchestration in October

Product, technology, and tools updates

Practitioner guides and how-tos

Customer and partner updates

Research, reports and deep-dives

September 2026

Product, technology, and tools updates

Practitioner guides and how-tos

Customer and partner updates

Research, reports and deep-dives

August 2026

Product, technology, and tools updates