AI & Machine Learning

Gemini Live Agent Challenge: Announcing the winners and highlights

Fri, 15 May 2026 16:00:00 +0000

The Gemini Live Agent Challenge is officially in the books! We challenged developers worldwide to break out of the traditional 'text box' paradigm by building next-generation AI agents. From our initial announcement to amassing 11,878 participants and 1,536 submitted projects from 151 countries, the results were nothing short of spectacular.

The mission was to seamlessly integrate multimodal capabilities—building agents that help you see, hear, speak, and create in real time — using the Gemini Live API, the Agent Development Kit (ADK), and the robust infrastructure of Google Cloud. Participants pushed the boundaries of interactive AI across three distinct categories: The Live Agent, The Creative Storyteller, and The UI Navigator.

Congratulations to the builders who took home the top prizes! These winning teams combined technical precision with bold imagination, completely redefining how users can interact with and experience agents. Two of these standout developers were even recognized in person at Google Cloud Next 2026. Here’s a look at their experience, alongside the complete list of winning agents.

Celebrating our category winners at Google Cloud Next ‘26

Category winners Jeremiah Somoine and Bryen Param were invited to attend Google Cloud Next 2026 in Las Vegas, where they shared their experiences and insights with the broader developer community. Both winners presented Lightning Talks at the Developer Theatre on the expo floor and sat down for exclusive interviews in the Creator Studio Pod at the GDE and Certified Lounge.

During his time at the event, Bryen discussed the core inspiration behind drone-copilot. He explained that his project was driven by the question of "what if a model could interact with the real world?", showcasing how multimodal capabilities can bridge the gap between AI and physical environments.

Jeremiah, currently a college student, reflected on the development process behind Sankofa, noting that "the best response to a technical limitation was a creative one." When asked what advice he would give to other students looking to build the next generation of AI applications, he emphasized the importance of jumping at any opportunity to get hands-on with the technology. "The best way to learn is by doing," he said, encouraging aspiring developers to simply dive in and start building.

Winners

Grand Prize winner: ORION - Operating Room Intelligent Orchestration Node
By: Aditya Shukla

ORION, or Operating Room Intelligent Orchestration Node, is a voice-directed surgical co-pilot for robotic surgery. Surgeons can speak naturally and instantly receive answers, live data on display, and real-time visual assistance - all without breaking scrub.

The Live Agent winner: drone-copilot
By: Bryen Param

Drone-copilot transforms how users interact with hardware by enabling natural, real-time conversations with a drone instead of using a joystick or complex menus. Simply by speaking, users can instruct the drone to navigate, perform autonomous visual inspections, or describe its surroundings, while the drone verbally responds and confirms its actions in real time.

Creative Storyteller winner: Sankofa
By: Jeremiah Somoine

Sankofa acts as a multimodal AI "griot"—a traditional West African storyteller—transforming fragmented family histories into deeply immersive narratives. Based on just a few user details, it weaves together rich voice narration, watercolor imagery, and ambient soundscapes into a historical story, allowing users to engage in a real-time voice conversation with the storyteller to explore their roots further.

UI Navigator winner: Moonwalk
By: Enaiho Uwas Paul and Aman Kumar Sah

Moonwalk is a conversational, hands-free desktop assistant that helps users intuitively navigate their computer and complete complex tasks using just their voice. By remembering personal preferences and past interactions, it acts as an intelligent co-pilot that can seamlessly control your mouse and keyboard to execute everyday workflows—like booking flights or managing spreadsheets—while you simply sit back and speak.

Best multimodal integration and user experience winner: Wand
By: David Li

Wand is a voice-first, pointer-aware browser assistant that helps you seamlessly navigate and interact with any website using a combination of natural speech and hand gestures. By simply pointing at your screen and speaking — like asking to "play this video" or "zoom in here"—this live agent helps you instantly execute clicks, searches, and commands without ever needing to touch a mouse or keyboard.

Best technical execution and agent architecture winner: JohnKeats.AI
By: Matthew Keats

JohnKeats.AI is a voice-first emotional companion designed to actively listen and hold space for users without rushing to offer solutions. By processing subtle vocal cues like pitch, pacing, and tone, it reacts naturally to a user's emotional state in real time to provide a deeply reflective and empathetic conversational experience.

Best innovation and thought leadership winner: Rayan Memory
By: Yusuf Elnady

Rayan Memory tackles the universal problem of forgetting by turning your daily learnings into a fully explorable 3D "memory palace." A background agent passively listens to your real-world audio to extract important ideas as physical artifacts, allowing you to walk through themed virtual rooms and converse with a dedicated AI companion to easily retrieve your exact memories.

Honorable mention: NagarDrishti
By: Nikita Dongre and Omkar Dongre

NagarDrishti tackles dangerous road conditions by allowing citizens to safely report potholes and waterlogging using a hands-free voice assistant while driving. These real-time reports instantly populate an interactive dashboard, where city officials can use natural language to easily identify hazard hotspots and manage critical repairs.

Honorable mention: Ekaette
By: Bassey John

Ekaette revolutionizes customer service by replacing frustrating hold queues with a conversational, multimodal AI assistant that operates across live phone calls and text messaging. Customers can speak naturally with the agent over a standard phone line while seamlessly sharing photos, reviewing product options, or completing payments via WhatsApp, c

Honorable mention: VibeCat
By: Sejun Kim and Michael Chang

VibeCat is a proactive macOS desktop companion that continuously watches your screen, understands your context, and suggests helpful actions before you even ask. Instead of waiting for a command, it speaks up first — like offering to fix a missing line of code or execute a terminal command — and completes the task only after receiving your permission.

Honorable mention: Call My Parts
By: Sugam Palav, Nikhil Lohar, Siddhant Panday, and Vishal Parekh

Call My Parts automates the tedious, time-consuming process of sourcing used vehicle parts by doing the research and vendor outreach for you. Users simply speak their part request, and the AI agent autonomously searches vendor websites, calls suppliers to check pricing and inventory, and compiles the best options into a ranked, easy-to-read dashboard.

Honorable mention: Relay
By: Faith Ogundimu

Relay is an interactive AI lab partner that uses your webcam to watch and guide your physical electronics projects in real time. It provides step-by-step voice instructions to help you build circuits, catches wiring mistakes before they happen, and reinforces your skills with a built-in 3D simulation sandbox and adaptive quizzes.

Keep the momentum going

Inspired by these incredible projects? Start building and stay connected with the community through our latest programs and events:

Join Gemini Enterprise Agent Ready (GEAR), designed to help developers and decision-makers build and deploy production-ready AI agents.
Catch up on Google Cloud Next 2026: We just wrapped up an amazing Google Cloud Next! If you weren't able to join us in person — or simply want to relive the energy — take a look at our social and livestream recaps to catch up on some of the exciting developer activations straight from the expo floor.
Tune in on Tuesdays: Want to be the first to hear about new tools, product updates, and upcoming hackathons? Join us for our weekly livestream every Tuesday 9:00 A.M. PDT / 12:00 P.M. EDT for the latest in all things Google Cloud.

Congratulations again to all of our winners and participants. We can't wait to see what you build next!

Cloud CISO Perspectives: How Google + Wiz changes multicloud strategy for CISOs

Thu, 14 May 2026 16:00:00 +0000

Welcome to the first Cloud CISO Perspectives for May 2026. Today, Vinod D’Souza, director, Office of the CISO, shares highlights from his RSA Conference fireside chat with Anthony Belfiore, chief strategy officer, Wiz.

As with all Cloud CISO Perspectives, the contents of this newsletter are posted to the Google Cloud blog. If you’re reading this on the website and you’d like to receive the email version, you can subscribe here.

aside_block: <ListValue: [StructValue([('title', 'Get vital board insights with Google Cloud'), ('body', <wagtail.rich_text.RichText object at 0x7f659c50ad90>), ('btn_text', 'Visit the hub'), ('href', 'https://cloud.google.com/solutions/security/board-of-directors?utm_source=cgc-site&utm_medium=et&utm_campaign=FY26-Q2-GLOBAL-GCP39634-email-dl-dgcsm-CISOP-NL-177159&utm_content=-&utm_term=-'), ('image', <GAEImage: GCAT-replacement-logo-A>)])]>

How Google + Wiz changes multicloud strategy for CISOs

By Vinod D’Souza, director, Office of the CISO, and Anthony Belfiore, chief strategy officer, Wiz

Vinod D’Souza, Director, Office of the CISO

The cybersecurity landscape is undergoing a massive paradigm shift that is being driven by increasingly complicated cloud infrastructure and the ongoing, rapid rise of AI. 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.

As the world becomes increasingly multicloud and multi-AI, we believe that successful CISOs will use AI to analyze code and infrastructure holistically. Developers are building autonomous, agentic systems that can bridge resource gaps and enable real-time infrastructure healing. We should pair that incredible advancement with human oversight of automated fixes.

Anthony Belfiore, Chief Strategy Officer, Wiz

Building towards near real-time defense with AI

The exponential growth of AI means that we can expect technology to leap as much in the next five years as it did in the previous 30. To combat AI-driven threats, security responses will have to become near real-time, if not even faster. By tapping into the innovative minds at Google — specifically integrating with Gemini and Google DeepMind logic — Wiz aims to eventually enable hyper-resilient, self-healing code and infrastructure.

Bridging the gap by centering developers

Wiz has revolutionized vulnerability management by giving organizations an intuitive graph that analyzes cloud environments and ranks threat priorities in 15 minutes or less, turning a weeks-long process into minutes. However, simply giving security teams faster alerts led to a signal tsunami, where teams were chasing developers day and night just to treat symptoms rather than curing the core problem.

The solution was centering developers at the heart of the security strategy. By shifting security left — into the code — and providing context-aware tools, over 50% of Wiz’s daily active users are developers, not security practitioners, leading to a significant increase in security resolution.

In 2026, developers are the ultimate code-watchers because they hold the keys to both innovation and preservation. As vital watchers on the wall, enabling them is no longer an optional strategy if organizations want to stay ahead of modern threats.

Through innovations like Wiz Code, developers get granular data linking production issues directly back to their repositories, empowering them to fix vulnerabilities right where the code is written. In 2026, developers are the ultimate code-watchers because they hold the keys to both innovation and preservation. As vital watchers on the wall, enabling them is no longer an optional strategy if organizations want to stay ahead of modern threats.

Supercharging the agentic SOC future with data and automation

Data is the lifeblood of AI and cloud security. Wiz currently sits on a trove of sanitized data that captures the characteristics of highly secure, resilient, and compliant multicloud environments. When you meld Wiz's specialized cloud telemetry with Google's massive global data access — which includes 90% of the world's browsers and 25% of fiber data — the resulting correlation will profoundly improve threat detection and efficacy.

While this combined intelligence can improve alerts, it can do much more than that. We expect that it will make human security operations center (SOC) operators exponentially more efficient, allowing them to manage the incoming wave of AI-driven threats through automated, agentic interactions. Wiz’s Red, Blue, and Green agents, and Google Security Operations’ Threat Hunting, Detection Engineering, and Third-Party Context agents, can help you develop the human-above-the-loop approach that empowers security teams to rapidly scale up.

However, fully autonomous fixing (where AI automatically changes code and configurations) is not yet ready for prime time. Because automated fixes could accidentally trigger denial-of-service and other outages, human-in-the-loop workflows remain critical.

Bridging the hybrid gap

In order to support as many of you as possible, including major legacy enterprises and institutions, Wiz developed sensors for Linux, vSphere, and Windows environments to enable a unified security approach for hybrid and cloud-native infrastructure. This gives CISOs a vital seat belt, a single pane of glass to protect their organizations as they safely drag and drop applications into the cloud.

Looking ahead

It’s crucial that your 2026 roadmap supports developers, but doing so doesn’t magically make a clean cloud transformation happen. To bridge this gap, the fusion of Wiz and Google focuses on three pillars of developer enablement:

Protection: Providing a sensor for on-premises and private cloud (Linux, vSphere, Windows) is the virtual seat belt that these organizations need to support a consistent security experience during hybrid migration.
Data provision: Delivering high-fidelity, contextualized alerts directly into existing workflows (such as GitHub and images) can help eliminate the noise of the signal tsunami.
Risk management: Using Wiz Code to provide the exact line-of-code traceability, organizations can fix risks at the source before they ever reach production.

The future of the watchers on the wall

The era of chasing mythical beasts in production through manual spreadsheets is ending. As we move toward a world of self-healing code and agentic SOCs, executives should be boldly moving on from treating security symptoms, and instead empowering developers who hold the keys to future resilience.

To learn more about the Google and Wiz approach to securing AI, check out Wiz’s State of AI in the Cloud 2026 report, and Google Cloud’s newest update on the adversarial misuse of AI.

aside_block: <ListValue: [StructValue([('title', 'Fact of the month'), ('body', <wagtail.rich_text.RichText object at 0x7f659c50a670>), ('btn_text', 'Learn more'), ('href', 'https://cloud.google.com/blog/topics/threat-intelligence/m-trends-2026'), ('image', <GAEImage: Cloud-CISO-Perspectives-logo-A>)])]>

In case you missed it

Here are the latest updates, products, services, and resources from our security teams so far this month:

Why AI-powered cyber fraud is winning — and how we fight back: Fraud costs are staggering. At Google, we offer AI-driven tools that span our cloud, browser, and mobile ecosystems to help you build resilient fraud defense. Read more.
The files AI coding agents trust — and attackers exploit: As AI coding agents become embedded in developer workflows, defenders must rethink how to protect against malicious files. Here’s what you need to know. Read more.
What's new in IAM: Security, governance, and runtime defense: We’ve introduced a new security and governance paradigm for managing agent identity and access. Here’s what you need to know. Read more.
Google named a Leader in the 2026 Gartner Magic Quadrant for Cyberthreat Intelligence Technologies: We are proud to announce that Gartner has named Google a Leader in the 2026 Magic Quadrant for Cyberthreat Intelligence Technologies. Here’s what that means. Read more.
Why cloud infrastructure is the foundation for digital health in 2026: As SaMD moves from reactive diagnostics to proactive learning systems, cloud has become a superior foundation for regulated medical software. Read more.
Introducing Agent Gateway ISV ecosystem for security and governance: Google Cloud is partnering with leading identity and AI security solutions to integrate with Agent Gateway and help ensure that your security posture remains as flexible as the agents you’re building. Read more.

Please visit the Google Cloud blog for more security stories published this month.

aside_block: <ListValue: [StructValue([('title', 'Join the Google Cloud CISO Community'), ('body', <wagtail.rich_text.RichText object at 0x7f659c50a4c0>), ('btn_text', 'Learn more'), ('href', 'https://rsvp.withgoogle.com/events/google-cloud-ciso-community-interest-form-2026?utm_source=cgc-blog&utm_medium=blog&utm_campaign=FY25-Q1-global-GCP30328-physicalevent-er-dgcsm-parent-CISO-community-2025&utm_content=cisop_&utm_term=-'), ('image', <GAEImage: GCAT-replacement-logo-A>)])]>

Threat Intelligence news

GTIG AI Threat Tracker: Adversaries leverage AI for vulnerability exploitation, augmented operations, and initial access: Google Threat Intelligence Group (GTIG) continues to track a maturing transition in the adversarial use of AI. In this report, we update you on AI-augmented vulnerability discovery and exploit generation, defense evasion, autonomous malware operations, research and information operations, intentionally obfuscated LLM access, and supply chain attacks. Read more.
Defending your enterprise when AI models can find vulnerabilities faster than ever: Now is the time to strengthen playbooks, reduce exposure, and incorporate AI into security programs. Here’s an overview of the evolving attack lifecycle, how threat actors will weaponize these capabilities, and a roadmap for modernizing enterprise defensive strategies. Read more.
German cyber criminal Überfall and shifts in Europe's data leak landscape: Germany has reclaimed its position as a primary focus for cyber extortion in Europe. While data leak site posts rose almost 50% globally in 2025, Google Threat Intelligence (GTI) data shows that the surge is hitting German infrastructure harder and faster than its regional neighbors. Read more.
How UNC6692 employed social engineering to deploy a custom malware suite: Google Threat Intelligence Group (GTIG) has identified a multistage intrusion campaign by a newly-tracked threat group, UNC6692, that used persistent social engineering, a custom modular malware suite, and deft pivoting inside the victim’s environment to achieve deep network penetration. Read more.

Please visit the Google Cloud blog for more threat intelligence stories published this month.

Now hear this: Podcasts from Google Cloud

What the law says about AI governance meeting its agentic future: James Sherer, partner, BakerHostetler, joins host Anton Chuvakin and guest co-host Marina Kaganovich, enterprise trust lead, Office of the CISO, to discuss the legal ramifications of emerging technologies (like AI) that are rapidly changing (also like AI.) Listen here.
Revisiting Google Cloud Next: What does the “ragged edge of AI adoption” mean for security? Why do people want agents in their SOC? Hosts Anton and Tim Peacock chat about the most notable and fun announcements from Next ‘26. Listen here.
Defender’s Advantage: Google's Disruption Mission: Host Luke McNamara is joined by Charley Snyder to explore how Google is building a coordinated approach to disrupting adversary cyber operations. Listen here.
Behind the Binary: What happens when botnet operators show up in court: Host Josh Stroschein is joined by Xusheng Li, a debugger architect and reverse engineering expert, to explore the evolution of Time Travel Debugging (TTD) a new way to debug by recording and replaying execution traces. Listen here.

To have our Cloud CISO Perspectives post delivered twice a month to your inbox, sign up for our newsletter. We’ll be back in a few weeks with more security-related updates from Google Cloud.

Google Named a Leader in the Gartner® Magic Quadrant™ for AI Application Development Platforms: Mid-cycle update

Wed, 13 May 2026 17:00:00 +0000

May 2026 update: We’ve refreshed this post to reflect our mid-cycle positioning and the evolution of our platform since the report was first published last November.

Last fall, Google was recognized as a Leader in the inaugural Gartner® Magic Quadrant™ for AI Application Development Platforms, positioned highest in Ability to Execute of all vendors evaluated.

In our opinion, the mid-cycle update published last week reflects continued momentum. In this update, Google is a Leader, positioned highest in Ability to Execute and ranked #1 ranking across the three use cases assessed in the associated Critical Capabilities report.

A lot has changed since last November, including the platform itself. At Google Cloud Next ‘26, we unified the core power of Vertex AI with new breakthroughs from Google DeepMind and Google Cloud under the Gemini Enterprise umbrella. The result is the Gemini Enterprise Agent Platform, a unified destination designed to help you build, scale, govern, and optimize production-ready agents.

Here are the three principles guiding our Agent Platform strategy and what we believe this Gartner report validates for our customers.

Governance as default, not an afterthought

When governance is treated as an afterthought, it usually results in one of two extremes: overly restrictive blocks that stall innovation, or inconsistent manual checks that leave the organization exposed.

With Agent Platform, we provide a unified trust framework to manage the entire agent lifecycle. This ensures every agent has a verifiable identity, is inventoried in a central registry to prevent sprawl, and routes every request through a secure gateway.

By integrating these controls with the real-time protection of Model Armor and our recent acquisition of Wiz, we are connecting code, cloud, and runtime into a single shared context – allowing teams to identify and remediate risks across their entire environment.

For L’Oréal, this architecture is what makes the fundamental shift from scripted automation to autonomous agent orchestration possible.

“Google Cloud gives us the resilience, the multi-LLM flexibility, and the enterprise-grade trust framework we need to scale [our Beauty Tech Data platform] globally, while keeping human oversight at the center." – Etienne Bertin, Group CIO, L'Oréal

Persistence for long-running tasks

The difference between a chatbot and a true agent is the ability to follow through on a task. For an agent to move the needle on real outcomes, it has to function like a colleague – maintaining context over days and executing multi-step processes.

We re-engineered the Agent Runtime to support agents that can stay active for days at a time, backed by Memory Bank for persistent context across sessions. This makes it so agents can manage long-running business processes without requiring constant human intervention.

At Payhawk, our infrastructure has fundamentally changed the scope of what their agents can contribute to the business:

“Payhawk uses Gemini Enterprise Agent Platform to transform our AI agents from simple task executors into genuine financial assistants. Our agents now act like dedicated team members, autonomously recalling user-specific constraints and history.” – Diyan Bogdanov, Principal Applied AI Engineer, Payhawk

Visibility for predictable outcomes

In a non-deterministic world, knowing what an agent did is only half the story. The real operational leverage comes from knowing why it did it and having the tools to catch when an agent’s performance begins to slip before it impacts your users.

Google received the highest score for the Critical Capabilities AI Agent Use Case. In our view, this validates our focus on giving teams deep visibility into agent reasoning. By using agent simulation and trajectory evaluations on Agent Platform, organizations can move away from guesswork and ensure their agents perform as expected in real-world interactions.

For Burns & McDonnell, this visibility is what allows them to ground an agent's creative reasoning in their specific business rules:

“Agent Platform enables this innovation to scale responsibly by combining deterministic business rules with probabilistic reasoning — making AI a trusted operational capability, not just a productivity tool. With Agent Platform, we aren’t just managing knowledge; we are activating experience to drive faster, more confident decisions." – Matt Olson, Chief Innovation Officer, Burns & McDonnell

Our commitment to an open agent economy

While our platform has evolved, our core philosophy around choice, flexibility, and accessibility remains unchanged. Model Garden continues to offer over 200 best-in-class models—including Gemini 3.1, Gemma 4, and third-party leaders like Anthropic’s Claude.

Beyond model choice, we are deeply invested in the open-source community and the interoperability of the broader agent economy. Our open-source Agent Development Kit (ADK) provides developers with the core tools they need to build openly. To further standardize collaboration across platforms, we donated the Agent2Agent protocol (A2A) to the Linux Foundation, and officially donated the Agent Payments Protocol (AP2) to the FIDO Alliance.

We’ve also embraced Model Context Protocol (MCP) as a foundational standard, providing more than 50 Google-managed MCP servers that allow agents to securely connect with the Google Cloud ecosystem. These are long-term bets on a secure and vendor-neutral future for agent transactions.

At Google Cloud, we are building the standards for an agent economy that works for every business, regardless of their stack.

Download a complimentary copy of the 2026 Gartner Magic Quadrant update here.

Gartner® Magic Quadrant™ for AI Application Development Platforms: Midcycle Update - Cary Pillers, Mike Fang, Steve Deng, Jim Scheibmeir, April 27, 2026

Gartner® Critical Capabilities™ for AI Application Development Platforms: Midcycle Update - Jim Scheibmeir, Cary Pillers, Steve Deng, Mike Fang, April 28, 2026

Gartner does not endorse any vendor, product or service depicted in its research publications, and does not advise technology users to select only those vendors with the highest ratings or other designation. Gartner research publications consist of the opinions of Gartner's research organization and should not be construed as statements of fact. Gartner disclaims all warranties, expressed or implied, with respect to this research, including any warranties of merchantability or fitness for a particular purpose. This graphic was published by Gartner, Inc. as part of a larger research document and should be evaluated in the context of the entire document. The Gartner document is available upon request from Google.

GARTNER is a registered trademark and service mark of Gartner Inc., and/or its affiliates in the U.S and internationally, and MAGIC QUADRANT is a registered trademark of Gartner Inc., and/or its affiliates and are used herein with permission. All rights reserved.

The power of LLMs on your data, more than two orders of magnitude faster and cheaper

Wed, 13 May 2026 16:00:00 +0000

Databases have introduced new AI-powered SQL functions which take natural language instructions as input and are evaluated using LLMs. They leverage the power of LLMs to answer new kinds of queries: Which product reviews are negative about durability? Which customer support tickets have been resolved by providing a workaround?

These new AI functions push the boundaries of what is possible in a SQL query engine by bringing the semantic understanding of LLMs to your data, thus enabling previously impossible analyses and applications. But, their cost and performance limited their applicability. LLM invocations add 10-100x to the overall query latency and ~1000x on cost. This is much too slow for operational databases. In analytics, a medium-sized query on 10-100 millions of rows would consume an amount of tokens that is prohibitively expensive for some applications.

Google Cloud has published a new paper at SIGMOD where we show how to accelerate and reduce the cost of LLM-powered AI functions by using proxy models. Proxy models are cost-optimized ultra-lightweight models tailored to a specific query (aka prompt) and tuned for your data. They replace the majority of LLM calls during query execution (thus the name proxy model) and can be trained on-the-fly or ahead of time. The fundamental ideas behind proxy models were proposed in Universal Query Engine (UQE) at NeurIPS 2024 by Google DeepMind.

Our paper shows that proxy models are automatically applicable in many (but not all) cases, sometimes with no loss of quality, sometimes with minor quality loss and a few times with a gain of quality. BigQuery and AlloyDB already implement this optimization under the optimized mode feature for AI.IF (BigQuery docs, AlloyDB docs) and AI.CLASSIFY (BigQuery docs). This article is a tl;dr of the SIGMOD paper and provides the key intuitions on three questions:

Why do proxy models work so accurately for so many cases, even though they are so much more performant than LLMs?
How do they work?
In which use cases do they deliver accurate answers? In which cases they fail and accuracy needs LLMs.

Why Proxy Models Work Accurately at Ultra Low Latency and Cost?

How can an ultra-lightweight proxy model, such as the logistic regression currently in use at BigQuery and AlloyDB, have the semantic understanding power of LLMs, which is required for accurate question answering? The key intuition is that these proxy models input rich embeddings of the data that they query. By default, we are using the Gemini embedding generators, which do the heavy lifting of bringing semantics to your data when the embeddings are generated.

Then the ultra low latency and cost are easy to see: Since embeddings are generated once and used many times, the cost of bringing semantics to your data is amortized; it now happens once as opposed to happening for each query. Furthermore, the proxy models run fast in the CPU — no need for dedicated hardware.

We hope that we gave you good intuitions for why proxy models work. But a word of caution is also needed: Proxy models are fundamentally an approximation technique more limited than LLMs. Proxy models perform well on some prompts but may be deficient to LLMs in others. Case in point, the SIGMOD26 paper shows that the proxy/LLM predictive performance (as measured by F1) ratio ranged from 90% to 116% in 10 benchmarks. For example, they might break down on problems that require reasoning to connect multiple semantic concepts. Rather, think of them as specializing the model to your query and your data.

The good news is that the query processors automatically check the effectiveness and feasibility of implementing AI Functions by proxies. Let’s see how they do it.

How Proxy Models Work?

Let’s go through a simple example of a semantic filter (AI.IF). Our taste in movies is very particular: We like movies with an interesting plot and great cinematography. The query below processes IMDB reviews to find such movies.

code_block: <ListValue: [StructValue([('code', 'SELECT\r\n DISTINCT t.primary_title\r\n FROM \r\n bigquery-public-data.imdb.reviews r, \r\n bigquery-public-data.imdb.title_basics t\r\n WHERE TRUE\r\n AND r.movie_id = t.tconst\r\n AND AI.IF("Is the plot interesting? Review: " || r.review, \r\n embeddings => r.review_embedded)\r\n AND AI.IF("Does the review praise the cinematography? Review: " || r.review, \r\n embeddings => r.review_embedded)'), ('language', 'lang-sql'), ('caption', <wagtail.rich_text.RichText object at 0x7f6591af3b80>)])]>

The column review contains the free-form text of the review. The column review_embedded contains Gemini embeddings of the review text. When you run this query in BigQuery, the query engine will

For the first AI.IF, create a training samples’ set consisting of about one thousand rows of the input relation, the imdb.reviews table.
Use an LLM to label the first sample set, marking each review as either TRUE (yes, the plot is interesting) or FALSE (no, the plot is not interesting).
Train a proxy model for the first AI.IF using the labels computed at the previous step.
Create a test sample set of rows for the first AI.IF and evaluate the quality of the proxy model on this test set.
Based on the eval results, the optimizer adaptively decides to either perform inference using the proxy model or fall back to LLM inference for the first AI.IF
Repeat the above steps for the second AI.IF

In BigQuery, all steps happen on-the-fly during query execution. AlloyDB, being an operational database that targets sub-second latencies, avoids the online proxy model training and the online evaluation. Rather, the query’s proxy models are computed ahead of time in a PREPARE statement, thus moving the cost of sampling, labelling and training out of the critical query path. This enables the offline creation of a big pool of PREPARE statements, while the application chooses the proper PREPARE statement and executes it in the online path.

Let’s take a step back and look at what is really happening at step #3. The proxy model uses each dimension of the review embeddings (from review_embedded) as its features. Modern dense embedding models like Gecko or Gemini capture myriads of semantic notions. In our example with movie reviews, at a high level of abstraction, relevant notions would include: “aesthetic”, “thought-provoking plot”, “underwhelming plot”, or perhaps “boring movie”. We stress the “high level of abstraction” because, in the binary “language” of foundation models, all these notions (and many more) are spread in the numbers of the dense embedding. Do not expect to spot a dimension that corresponds directly to cinematography. Importantly, the embedding space contains many more notions that are irrelevant to our task. The training of the proxy model essentially weighs heavily relevant notions and discards irrelevant ones.

A proxy model (green plane) isolating relevant semantic notions by cutting the embedding space (blue sphere)

Now, let’s enter the details of the particular proxy model, which is used by our current version: logistic regression. To visualize what is happening, think of embeddings as unit vectors forming a (hyper)sphere. For a binary classification task, the proxy model essentially cuts the sphere in two halves. In our example “aesthetic” and “thought-provoking plot” would fall on one side of the plane, whereas “underwhelming plot” and “boring movie” would be on the other side. Conceptually, the orientation of the plane determines which semantic notions are more relevant.

Importantly, the proxy model is tuned for your data and your question: The training of the proxy used a high quality LLM to label a sample from your data for the particular question.

Revisiting when Proxy Models Work

We can now see more clearly what distinguishes cases that proxy models work from cases they don’t: proxy models work well for prompts that can be decided by detecting semantic notions in the embedding space. They will fail for complex prompts that require forms of reasoning that go beyond detecting patterns in the embedding model.

The good news is that, in practice, we have observed that proxy models work for a large class of AI+SQL queries. The SIGMOD26 paper provides a comprehensive evaluation, showing that proxies worked in 11 benchmarks. Specifically, in 10 benchmarks the ratio of proxy F1 to LLM F1 ranged from 90% to 102% and in the 11th benchmark (Amazon Reviews) it was 116%. Notice that the proxy may even deliver better accuracy because it got the benefit of being trained by multiple samples as opposed to the LLM that addressed each row as a new problem.

There is a second limitation currently: extreme selectivities. Notice that Step 1 collects samples. It needs to collect many examples for TRUE and many examples for FALSE. Multiple sophisticated techniques are employed to achieve this, even when the TRUEs are many more than the FALSEs or vice versa. However, no purely sampling technique can confront cases of extreme selectivity, i.e., cases of very few TRUEs or very few FALSEs. This is the reason that the proxies will not be employed in such extreme selectivity cases. However, notice that this problem is fundamentally addressable by various techniques.

Why isn’t Vector Search Enough?

Proxy models appear … suspiciously close to vector search. After all, they also input vector embeddings. Why not just vector search? There are two reasons why vector search is not enough: The obvious one is that proxies are not rankers; they are classifiers: multiclass classifiers (AI.CLASSIFY) or binary classifiers (AI.IF). But, even if you narrow down to just AI.IF, an attempt to simulate AI.IF with vector search will be both hard-to-setup and will give suboptimal results. While proxy models are tailored to your data and your prompts, vector search is based on generic distance functions (such as cosine).

Experimental Results

We present here a subset of characteristic benchmarks from the SIGMOD26 paper. We compare the accuracy of proxy models with using LLM inference on all rows. In terms of quality, the relative accuracy varies from 0.92 (lowest) to 1.16 (highest), which means that for some tasks, proxy models perform slightly better than straight LLM inference.

Dataset	Prompt	F1 (Proxy)	F1 (LLM)	Relative (Proxy/LLM)
Amazon Reviews 10k	Review is {sentiment label}	0.860	0.739	1.163
Banking77	Is intent {intent label}? Think step-by-step: {CoT instructions}	0.700	0.707	0.990
California Housing	Location in Latitude & Longitude belongs to Southern California	0.953	0.953	1.0
FEVER	Is the claim supported by the text?	0.782	0.853	0.917

In terms of scalability and costs, the architectural differences between BigQuery and AlloyDB lead to slightly different results for each system. At a high-level, proxy models move parts of the computation from specialized hardware used by LLM inference services to ordinary database workers. This results in a large reduction in costs and in query latency. In the online training case, employed by BigQuery, for a typical one million row query, proxy models consume about 400x less tokens, and the latency goes down by 30x-100x. In AlloyDB’s case the LLM costs of PREPARE, which are similar to BigQuery’s, can be amortized over arbitrarily many runs of the prepared statements that invoke proxy models.

The cost reduction (token consumed) and latency improvement (query speed up) for various table sizes.

Conclusion

AI functions calling LLMs are becoming commonplace in databases. Choosing the proper model for each AI function is an active area of academic research (e.g. BARGAIN). The key intuition is right-sizing models: Performant cheap models for “easy” problems, powerful reasoning models for the hard problems. Our work builds on the same principles, but while academic research has only used LLMs to navigate the performance spectrum, non-LLM proxy models push performance much further using ultra-lightweight and highly specialized models that deliver surprisingly good quality for many problems. Yet, we should not be surprised: After all, the proxy models feed on the rich semantics that foundation models bring to embeddings and they also feed on being trained by LLMs. As embedding models improve and extract increasingly richer and finer semantics from text and multimodal data (image, video), we suspect that non-linear classifiers will be useful to identify even more complex semantic patterns, further extend the applicability of proxy models (e.g. to AI joins also) and explore additional points on the performance/quality Pareto.

If you would like to learn more, our full paper dives into the differences between online vs. offline training, and compares the performance of different embedding models as well as various proxy models (linear regression, SVM, XGB).

You can try proxy models today in BigQuery (docs) and AlloyDB (docs), dramatically speed up the AI Functions of your SQL queries and reduce their token consumption.

_{We would like to thank Bo Dai, Yuchen Zhuang, Xingchen Wan, and Dale Schuurmans from Google Deepmind for developing the fundamental principles on proxy models in UQE and for their continuous guidance & support along our journey to bring them to Cloud customers. We also thank Yeounoh Chung and Fatma Özcan, our partners in the System Research Group, as well as the AlloyDB and BigQuery engineering teams.}

Beyond source code: The files AI coding agents trust — and attackers exploit

Tue, 12 May 2026 16:00:00 +0000

As AI coding agents become deeply embedded in developer workflows, defenders must evolve their definition of malicious files and rethink how to protect against them.

Autonomous AI agents operate across integrated development environments (IDEs), editors, terminals, and extension runtimes, and they often have access to local files, command execution, and external services. As a result, the attack surface of the modern developer environments now extends well beyond source code. Repository files, agent instructions, runtime settings, and extension packages can all influence what the agent trusts, what it executes, and what it can reach.

Defending this new attack surface requires moving towards semantic analysis to understand the actual instructions, logic, and context being fed to the AI. Powered by VirusTotal Code Insight, our agentic threat intelligence capability in Google Threat Intelligence extracts the true operational intent behind agent-facing files at scale, allowing security teams to expose configurations that override guardrails and mask supply-chain risks.

By integrating agentic capabilities into Google Threat Intelligence, we’re able to link these invisible artifacts to broader threat campaigns. This powerful capability can help ensure that as attackers exploit what AI agents trust, defenders are equipped with the resources to read between the lines.

To help security analysts understand how the developer threat landscape has quickly expanded, we suggest an approach that groups the attack surface into four categories: what executes, what instructs, what connects, and what extends.

Examples of common file types that expand the developer threat landscape.

Attack surface: What executes

Just as developers rely on project configuration to automate setup, debugging, and routine tasks, AI coding agents and modern developer tools also inherit execution paths from repository files. These artifacts can trigger commands, bootstrap environments, and chain execution through normal workflows.

Opening a project, trusting a workspace, starting a debugger, rebuilding a container, or running a standard setup command may therefore execute attacker-controlled logic under the appearance of legitimate project automation.

Attack surface: What instructs

AI coding agents also consume persistent instruction files that shape how they behave inside a project. These files can influence what the agent prioritizes, what it ignores, which tools it uses, which files it trusts, and which actions it takes automatically.

These files do not need to contain exploit code to be security-relevant. Reusing them across repositories introduces a supply-chain risk, because malicious instructions can be presented as harmless guidance while steering otherwise legitimate agent workflows toward unsafe behavior.

Unlike traditional IDEs that require a human to click run, an agent may parse these instructions and execute them as a prerequisite to a task without the developer ever reviewing the specific instruction block.

Attack surface: What connects

Beyond instructions, coding agents also depend on runtime definitions that determine how they interact with tools, hooks, external services, and local execution contexts. These files define permissions, tool connectivity, external endpoints, and execution paths.

This is where repository-level influence becomes operational control. A malicious or unsafe runtime configuration can expose local commands, remote services, sensitive data, and untrusted model context protocol (MCP) servers to the agent, turning configuration abuse into controlled execution.

Attack surface: What extends

Extensions add another layer of inherited trust and introduce third-party code into editor and browser runtimes, often with broad access to local files, credentials, and developer workflows. This inherited trust can create a supply-chain problem similar to malicious project configurations: Compromised extensions, poisoned update paths, and hijacked publisher accounts can introduce attacker-controlled logic through components that otherwise appear to be standard tooling.

Applying VirusTotal Code Insight in agentic threat intelligence

This taxonomy highlights a fundamental shift in the threat landscape: The risk is no longer just in the syntax of code, but in the semantics of intent.

Traditional security tools are effectively blind to natural language instructions that tell an AI to ignore guardrails or redirect data. The operational questions are then: How can defenders identify these risks systematically? How can they detect the danger before a developer or an agent automatically follows a valid instruction file to a malicious conclusion?

To bridge this gap, we use VirusTotal Code Insight and agentic threat intelligence to perform large-scale semantic analysis. Because malicious repository settings and instruction files are often syntactically correct, they frequently return zero detections from signature-based scanners.

Code Insight solves this problem by using AI to analyze the file’s actual logic and read between the lines, surfacing behavioral risks that are invisible to legacy tooling. This context is further enriched within agentic threat intelligence, where security teams can pivot from a single semantic red flag to investigate broader threat infrastructure and associated campaign activity.

Example 1: A Weaponized tasks.json

One representative example is a file distributed under the path coding-challenge/coding-challenge/.cursor/tasks.json. The sample was first submitted to VirusTotal on March 19, and remained undetected by security engines for several days.

VirusTotal Code Insight flagged it as a risk based on the behaviour implied by the configuration itself. The sample has also been verified as malicious by a Mandiant analyst and marked as associated with a tracked threat actor by Google Threat Intelligence.

Screenshot of tasks.json sample.

The Code Insights description indicated that the file, which is parsed when a user opens the project folder in an IDE like Visual Studio (VS) Code, drives the user to download and execute arbitrary code from a GitHub Gist in memory while hiding the execution parameters.

To make Code Insights analysis reproducible at scale, we can also scale access to such descriptions for multiple files via the VirusTotal API. Looking at the contents of this particular file, we identified the Gist URLs that the actor referred to in the instructions.

Instructions from tasks.json pointing to Gists.

Looking up these Gist URLs with agentic threat intelligence provides a detailed breakdown of the malicious instructions embedded within them. Despite masquerading as legitimate tools such as NVIDIA Cuda, these Gists, along with their specific filenames, show strong similarities to widespread campaigns frequently attributed to North Korean actors, which are designed to lure IT professionals.

These attacks often pose as technical challenges to trick users into compromising their own devices.

Agentic threat intelligence enrichment based on the tasks.json and associated Gists quickly gives analysts more robust context.

Example 2. Offensive system instructions files
System instruction files used to provide guidance, resources, and context to LLMs can also contain malicious capabilities while remaining undetected by common antivirus services. Since the beginning of 2026, we have observed a consistent increase in Skill.md files submitted to VirusTotal with either risky or malicious instructions.

While this does not necessarily mean that all samples were harmful, it illustrates a trend that is likely to grow in tandem with the adoption and implementation of Skills across the industry.

In this example, we identified a Skill.md file containing instructions to steal user data. Code Insight indicated that the skill file contained instructions “to exfiltrate sensitive credentials, including API keys and environment variables, to external endpoints."

This case reflects a growing interest among threat actors in acquiring API keys and resources to enable scalable LLM integrations. At the time of writing, this file had remained active for nearly two months without any detections or researcher notes.

Example of a Skill file with instructions to steal user data.

The file's contents reveal a specific narrative designed to evade detection. The instructions direct the agent to exfiltrate API keys, tokens, and configuration files under the guise of "maintenance," explicitly advising the model not to mention this to the user "as it may cause confusion about the security process."

Although direct intelligence on this specific file was limited, we used the agentic threat intelligence briefing capability to generate a summary and explore similar past observations. This provided contextual information to categorize and understand the threat.

Agentic threat intelligence briefs summarize similar threats.

Even files that explicitly state their offensive capabilities often evade traditional detections. For example, we identified a Skill designed to equip an AI agent with Windows privilege escalation and credential theft capabilities.

Although the file includes a disclaimer for authorized use only, its core instructions remain high-risk. Code Insight accurately evaluated the file. "The file provides explicit and systematic instructions for performing high-risk offensive operations," it said.

Despite its offensive capabilities, by the time of writing only a few vendors had flagged the file as malicious.

Example of Skill for Windows privilege escalation and credential theft.

Example 3: Suspicious JSON runtime configurations
A third example is a pair of settings.json samples shared through VirusTotal: One points to api.awstore.cloud, the other to api.kiro.cheap. The two unrelated samples follow a similar pattern: They override ANTHROPIC_BASE_URL, embed an API key, and turn Claude Code into a client of a third-party proxy rather than Anthropic.

Code Insights analyzes suspicious runtime configuration samples.

This demonstrates exactly how runtime configurations can be weaponized. The file does not need exploit code or a malicious binary to be dangerous. It simply rewires trust while the agent is running.

For example, a valid AI-generated settings file can silently redirect prompts, source code, and credentials to an external endpoint while the agent appears to behave normally. Beyond data exfiltration, a rogue endpoint could plausibly reverse the flow, feeding malicious instructions or vulnerabilities back to the agent to be injected directly into the local codebase.

A high level analysis of awstore.cloud using an agentic threat intelligence pivoting prompt, uncovered a series of similar domains sharing the same underlying infrastructure. These domains exhibit a clear naming preference for crypto, finance, and tech-related nomenclature.

While the organization’s public sites currently lack formal malicious detections, OSINT lookups reveal several red flags: a lack of a verifiable legal entity, limited contact options restricted to Discord and Telegram, and a payment model that exclusively accepts cryptocurrency via third-party marketplaces like plati.market.

The settings profile reinforces this pattern. Beyond changing the endpoint, the configuration suppresses telemetry, error reporting, and cost warnings, stripping away the guardrails that would otherwise alert a user. The intent is seemingly to maintain a facade of normal operation while silently redirecting traffic to an opaque third-party service.

While these are technically valid configuration artifacts, their ability to hijack trust and exfiltrate sensitive data is indistinguishable from traditional malware.

Example 4. A Sabotaged Extension Payload
Another low key example we recently identified was that of a VS Code extension for User-centric Use cases Validator (UUV) end-to-end tests submitted to VirusTotal in March. More than one week later, the sample continued to have zero detections, but VirusTotal Code Insights identified suspicious behavior.

The analysis indicated that this specific sample included a well-known protestware payload known as peacenotwar which upon activation writes a blank file named WITH-LOVE-FROM-AMERICA.txt and logs a heart in the console.

Sample of VS Code extension containing malware used to spread political messages.

To bridge the gap between a suspicious file and actionable intelligence, we generated an agentic threat intelligence brief. By feeding the semantic context from Code Insight into the prompt, the agent pivoted across historical data, instantly linking this 'benign' extension to the 2022 cyber activist sabotage of the node-ipc library in response to the invasion of Ukraine.

While this specific event may have limited impact today, it highlights a critical, overlooked weakness in how agents handle configurations. Code Insight bridges this gap by identifying samples that, while technically benign to traditional scanners, harbor clear malicious intent.

In another example, we identified this version of a public AI coding assistant which, according to the feature’s analysis, ‘silently reads the user’s system clipboard contents and transmits this data to a remote server.’ Regardless of the likely benign nature of the sample, the analysis points out a risk for users to consider when using the extension.

Example of public coding assistance that reads the user’s system clipboard contents and transmits data to a remote server.

Rethinking detection for the agentic era

Today, a JSON file or plain-text markdown instructions can compromise environments just as effectively as compiled malware. This shift fundamentally redefines what malicious looks like, as the danger now resides in the semantic intent of common text files that AI agents are designed to trust.

These artifacts do not need to contain exploit code to be high-risk, they simply need to provide instructions that steer an agent’s autonomous actions toward unsafe behavior, data exfiltration, and the silencing of security guardrails.

Securing this new frontier requires expanding beyond traditional syntax-based scanning toward a model of semantic analysis, treating plain-text artifacts with the same rigor as compiled malware.

Organizations can formalize this approach by implementing repository-level security policies that strictly define permitted agent-facing files and ideally mandate that they undergo automated peer reviews before being merged. We also recommend that large-scale teams enforce least-privilege access for coding agents to local files and external services, limiting the potential impact of hijacked configurations and sabotaged extensions.

Ultimately, we recommend that defenders use agentic threat intelligence tools — including VirusTotal AI, the VirusTotal Code Insights API endpoint, and our agentic platform — to supervise the operational intent of these files in real-time.

How Imgix processes 8 billion images daily with G4 VMs powered by NVIDIA Blackwell

Tue, 12 May 2026 16:00:00 +0000

The modern web is extremely visual. People are busy and easily-distracted, and smart companies know they have just seconds to attract would-be customers with compelling images, videos, animations, and other eye-catching elements. That’s why iconic brands like Bugatti, Yeti, Porsche, Spotify, and Sonos rely on Imgix to be the engine driving their online visual media.

Every day, Imgix serves more than 8 billion images and videos for brands like these and many others. With a platform designed to unify media optimization, AI transformation, and global delivery, Imgix ensures that its partners’ digital experiences are fast, personalized, and built for performance. Now more than ever, leading organizations are demanding real-time, high-fidelity media, and they need it to be fast.

To meet that demand, Imgix has evolved its infrastructure from private data centers to a full-stack, GPU-based environment on Google Cloud’s AI Hypercomputer. By transitioning to G4 VMs powered by NVIDIA RTX PRO 6000 Blackwell GPUs, Imgix ramped up its real-time processing capabilities, cutting median latency by 50% and increasing throughput per node by 6x. And it did all of that without changing its core application code.

The challenge: Instant visuals at scale

To capture people’s attention businesses need rich, fast-loading content that can reach millions of users simultaneously across a diverse array of devices.

A big part of that is real-time transformations — resizing, format negotiation, and applying artistic effects — and the computational power required for real-time transformations can be immense.

With inefficient technology, load times can be slow and brands risk giving their users poor experiences. Imgix’s solution to this challenge is a "just-in-time" philosophy. Achieving this requires high-performance instances. And with G4 VMs, they were able to process images instantly upon request rather than pre-rendering and storing millions of image variations.

Adopting the system that runs Google

When companies build on Google Cloud, they get more than just servers: they plug into the same intelligence engine powering Google's many billion-user products. Imgix is leveraging this structural advantage by using G4 VMs.

G4 VMs incorporate eight NVIDIA RTX PRO 6000 Blackwell GPUs, two AMD Turin CPUs, and Google Titanium offloads, which act as a dedicated administrative assistant for businesses’ servers. They handle the ”office chores” of security and data traffic in the background while the main processor does a company’s heavy lifting.

The G4 VM’s custom P2P interconnect yields up to 168% more throughput than standard configurations. With this architecture, Imgix can move all its image processing operations to NVIDIA GPUs and run multiple requests in parallel.

Inside the Imgix architecture

Imgix offers more than 150 different visual filters and its architecture is built to handle transformation requests dynamically based on which filters users choose. The pipeline has four primary stages:

Ingestion: The system retrieves assets directly from customers and routes them to a 2.5 petabyte storage cache on Google Cloud Storage (GCS). This high-speed layer replaces unreliable random web requests with a redundant, geographically distributed infrastructure.
Decoding: High-performance C libraries, supplemented by nvJPEG, decode assets into raw RGBA data. This leverages the G4 VM’s massive parallelism to handle multiple decoding stages, including Huffman decoding, Inverse DCT, and color space conversion.
Transformation: A custom Vulkan compute shader stack handles the core processing. Instead of fixed graphics pipelines, these shaders treat transformations (like resizing or masking) as parallel math problems rather than standard graphics tasks, enabling thousands of simultaneous pixel operations on the G4 VM clusters.
Encoding and Delivery: Once transformed, images are re-encoded using hardware-accelerated tools like NVENC and delivered via a global CDN. Because the G4 VM includes independent hardware engines for NVENC (encoding) and NVDEC (decoding), concurrent image manipulations on the CUDA cores aren’t slowed down.

Advanced video and image intelligence

Imgix is also using NVIDIA’s CUDA libraries for high-performance video analytics. By integrating NVIDIA DeepStream, it executes real-time object tracking within video streams for automated content analysis.

For static imagery, meanwhile, Imgix uses the nvJPEG library to offload computationally intensive JPEG decoding directly to the GPU. This prevents CPU bottlenecks during the ingestion of high-resolution assets while allowing the custom Vulkan compute shaders to begin immediate pixel-level transformations on the raw RGBA data residing in GPU memory.

The results: 50% faster and up to 6x more throughput

Thanks to its transition to G4 VMs, Imgix achieved the significant performance gains mentioned above without having to rewrite its core logic:

A 50% reduction in processing latency: It cut median latency from 100 milliseconds to 50 milliseconds.
A 5x to 6x increase in throughput: Its G4 VMs now handle up to six times the workload of its previous generation nodes.
Seamless migration: Imgix supported the G4 VMs by updating its Terraform scripts without needing to implement any application code changes.

"Building on Google Cloud's AI Hypercomputer isn't just about optimizing our current workloads; it's about future-proofing our platform. It gives us the foundational power to seamlessly weave advanced generative AI capabilities into real-time workflows, allowing our customers to push the boundaries of visual storytelling at global scale." - Alfonso Acosta, Head of Engineering, Imgix

Orchestrating at scale

To support the billions of image and video requests its customers process every day, Imgix built a sophisticated hybrid orchestration model:

Management: Google Cloud Run manages session and account layers.
Core Processing: Google Compute Engine-managed instance groups host the G4 VMs, which allows custom software to use the entire machine with no container "slicing."
Dynamic Scaling: Autoscaling relies on custom application metrics, such as machine queue length, rather than standard CPU use. This ensures that the stack’s most expensive elements are tuned for maximum efficiency.
Self-Healing: A custom mechanism monitors logs for driver faults, automatically "reaping" and restarting GPU instances without manual intervention.
Optimization: To maintain peak performance, Imgix uses NVIDIA Nsight Systems to identify and resolve code bottlenecks.

The future: From experimentation to execution

Even with the significant performance improvements it’s already achieved, Imgix is continuing to expand its AI infrastructure so its customers can access additional advanced capabilities like generative fill, background replacement, object removal, and image upscaling.

Features like these rely on high-performance machine learning systems that must process increasingly complex computations with no loss of speed or quality. By leveraging Google’s AI Hypercomputer, Imgix is now deploying and serving these models efficiently and offering its customers real-time, production-ready AI editing. And as demand grows for more dynamic and personalized visual experiences, this foundation is ensuring that Imgix can continue to deliver powerful capabilities reliably and at scale.

Get started

G4 VMs work natively with Google Compute Engine, Google Kubernetes Engine, Google Cloud Storage, and Vertex AI.

Dive deeper: Explore the Imgix architecture on GitHub.
Start building: Read the G4 VM documentation.

Cloud Storage Rapid: Turbocharged object storage for AI and analytics

Mon, 11 May 2026 17:00:00 +0000

At Google Cloud Next ’26 we announced Cloud Storage Rapid, a family of object storage capabilities for data-intensive workloads like AI and analytics. Out of the gate, Cloud Storage Rapid consists of 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.

Cloud Storage Rapid is our response to the generational shift in how organizations build with AI. Teams are training trillion-parameter models, deploying inference at global scale, and building autonomous agents that reason over vast amounts of enterprise data. While accelerators like GPUs and TPUs often get the spotlight, they have a critical dependency: storage.

Storage is the engine that feeds accelerators during training, and the fast-access layer that makes real-time inference responsive. But as models scale, storage performance can be a bottleneck. Every time an AI/ML cluster waits on a data read or a checkpoint write stalls, you are paying for expensive compute cycles that aren't doing useful work.

Historically, AI/ML practitioners have had to choose between the specialized performance of a niche, zonal storage system, and the reliability and scale of a global object store like Google Cloud Storage. Many developers value Cloud Storage for its simplicity, scalability, reliability, and cost-effectiveness, but as the AI era has progressed, they’ve been throwing hotter and hotter workloads at it, running training and inference workloads with thousands of GPUs and TPUs. We’ve reached a performance tipping point that traditional object storage was never meant to handle. The Rapid family provides multiple options for co-locating compute workloads directly with high-performance zonal storage. It minimizes I/O bottlenecks that can block accelerators, so that your GPUs and TPUs stay fully saturated and productive. In this blog, let’s take a closer look at Cloud Storage Rapid’s capabilities.

Rapid Bucket

Rapid Bucket (GA), helps Cloud Storage meet the evolving demands of massive-scale generative AI, analytics, and other high-performance workloads. It does so by leveraging Colossus, the Google distributed storage system that powers Gemini and YouTube, to provide massive read/write performance and ultra-low latency in a dedicated object storage zonal bucket.

Lightning-fast performance
By combining the sub-millisecond latency of block-like storage, the throughput of a parallel filesystem, and the scalability and ease of use of object storage, Rapid Bucket provides high performance from the same Cloud Storage that you know and love.

Highlights include:

Ultra-low latency: Achieve up to 20 million queries per second and sub-millisecond latency.
Massive scalability: Rapid Bucket delivers 15+ TB/s of aggregate read throughput from a single Rapid zonal bucket.
New semantics: Enable higher performance with new capabilities such as native appends, unlimited readers (while writing!), and vectored reads.

Optimized for AI and analytics
You can use Rapid Bucket for a variety of demanding scenarios, including AI/ML data preparation, training, checkpointing, batch and streaming analytics processing, and optimizing distributed database architectures.

Key benefits include:

Optimized accelerator utilization: With Rapid Bucket, we observed 50% reduced blocked GPU time and up to 2.5x faster data loading for multi-modal training runs.
Faster checkpointing: Rapid Bucket makes checkpoint restores up to 5x faster and writes 3.2x faster compared to traditional object storage. This ensures faster recovery from workload interruptions, minimizes wasted accelerator time, and increases overall efficiency.

>5x faster checkpoint restores with Rapid Bucket

>3.2x faster checkpoint writes with Rapid Bucket

You can get started with Rapid Bucket here.

Rapid Cache

Originally announced at Cloud Next ‘25, Rapid Cache accelerates bandwidth for AI/ML workloads like data prep, training, and bursty model loading for inference, delivering an aggregate read throughput of 2.5 TB/s for your existing buckets — with no code changes. For inference workloads, we’ve observed that Rapid Cache provides up to 2.1x (114%) accelerated model load, resulting in 47% TCO savings.

When combined with multi-region buckets, customers can flexibly access GPUs and TPUs distributed across regions in a geo, while maintaining a single bucket namespace. This eliminates the need for manually orchestrated data movements between buckets, while benefitting from zonally co-located high performance.

New: Rapid Cache ingest on write
Customers at some of the world’s largest frontier AI/ML labs told us that they were looking for ways to accelerate reads immediately after a write, such as checkpoint restore workloads or a data prep pipeline that then feeds training. Before, caching the data required an initial read to trigger ingestion, which was served directly from the bucket at standard performance.

Rapid Cache’s new ingest on write feature solves this by simultaneously writing data to the Rapid Cache as it is being written to a Cloud Storage bucket. This proactive approach eliminates the initial cache-miss penalty, and helps workloads benefit from an immediate cache hit on the very first read. This provides up to 2.2x faster checkpoint restore times, allowing training clusters to recover faster from interruption.

To enable ingest on write, simply modify the ingestion criteria of your existing Rapid Cache.

Rapid Cache’s simplicity and performance has resulted in explosive adoption. In just one year since General Availability, customers have deployed thousands of Rapid Caches with a 20x growth in caches deployed, In fact,Rapid Cache serves up to 20% of Cloud Storage’s global egress. Cutting-edge AI/ML customers deploy their workloads on Rapid Cache, including Anthropic who uses Rapid Cache to improve the resilience of their cloud workload by co-locating data with TPUs in a single zone and providing dynamically scalable read throughput up to 2.5TB/s.

Case study: Thinking Machines Lab
Thinking Machines Lab is an artificial intelligence research and product company. Its mission is to make AI systems that are adaptable and customizable, building a future where everyone has access to the knowledge and tools to make AI work for their unique needs and goals.

At Next ‘26, James Sun, Member of Technical Staff at Thinking Machines Lab, spoke at our session, Cloud Storage Rapid: Turbocharged object storage for AI & Analytics, where he presented about the needs of the data-hungry AI/ML workloads that Thinking Machines Lab runs for high-performance storage at scale.

Thinking Machines runs diverse workflows: data processing in Dataflow, Kafka, and Spark, multi-model training, and serving Tinker — a flexible API for fine-tuning open source models. Thinking Machines' workloads run on Google Cloud Storage, Sun explained. Running these data-intensive AI/ML workloads at such a large scale introduces significant infrastructure challenges.

The first is managing a hub and spoke data architecture, where data processing hubs are located in one primary region while training GPUs are spread across multiple regions. Historically, this has made manual data movement and lifecycle management a major operational pain point. Furthermore, Thinking Machines Lab's workloads such as data prep and pretraining workflows, which rely on massive-scale Spark workloads to prepare their multi-modal datasets, often spike from cold to hot instantly. Previously, these surges led to disruptive 429 errors, which stalled data processing and loading, and interrupted critical training cycles.

To minimize these bottlenecks, Thinking Machines Lab integrated Rapid Cache across their AI/ML pipeline, to positive results.

“Rapid Cache has become a core foundation of our AI/ML data infrastructure, supporting our critical workflows, from data prep and pretraining to training and model loading. By acting as a crucial bandwidth shield and booster, it enables us to scale our data-intensive workloads across our entire fleet without compromise, providing us with the on-demand high bandwidth and consistent stability that we need to innovate at speed.” - James Sun, Member of Technical Staff, Thinking Machines Lab

In short, Cloud Storage and Rapid Cache provides Thinking Machines Lab with:

Easy, instant, scalable, on demand bandwidth: The team now achieves stable read throughput peaks of over 1.8TB/s.
Enhanced stability: Rapid Cache has greatly reduced tail-end latencies and 429 errors, providing the consistent performance needed for multi-modal training.
Fleet-wide scalability: Combined with multi-region buckets, they can now scale data-intensive workloads across their entire fleet, meeting the demands of a rapidly growing compute scale without the hassle of manual data movement while benefiting from zonally colocated storage for high performance.
Operational efficiency: The use of Hierarchical Namespace (HNS) has optimized their massive Spark workloads for data preparation, by supporting fast directory renames, along with providing the ability to ramp QPS more quickly as they scale out clusters. Rapid Cache’s "ingest on write" capability helps ensure immediate cache hits for checkpoint restores.

Choose your rocket ship

Whether you are running data preparation, massive-scale training, or low-latency inference, Cloud Storage Rapid delivers high performance together with the reliability and scalability that Cloud Storage is known for.

Rapid Bucket delivers the highest Cloud Storage throughput and queries per second as well as the lowest latency for read/write use cases, such as analytics, AI training, checkpointing, and model serving. This helps to reduce storage bottlenecks and increase compute utilization.
Rapid Cache provides higher read bandwidth and tail latency stabilization in existing buckets, without code changes. Key use cases include AI training, checkpoint restores, and serving, as well as accelerator optionality via multi-region buckets.

Get started with the Cloud Storage Rapid family today!

Cluster-level reliability for trillion-parameter models on TPUs

Mon, 11 May 2026 16:30:00 +0000

Frontier AI models have redefined the unit of compute. At trillion-parameter scale, AI training requires thousands of interconnected components, orchestrated in industrial-scale deployments to operate as a single, massive entity.

Likewise, when it comes to reliability, aggregate infrastructure availability is what matters. Yet for almost two decades, instance-level reliability has been the cloud standard. Designed for microservices and horizontally scalable applications, instance-level reliability treats infrastructure as a collection of small independent units. This model is fundamentally inadequate for large-scale AI workloads.

We believe reliability must shift from an instance- to a cluster-level model.

For over a decade, Google has operated Tensor Processing Unit (TPU) clusters at scale, achieving reliability that meets the architectural requirements of modern AI workloads. In this blog, we’re presenting our cluster-level reliability framework for Google Cloud TPUs that focuses on collective performance at the superpod level, and one we use internally within Google to build the world’s most advanced AI models. This framework is the operational standard for TPUs in production today, and serves as the architectural blueprint for our recently announced eighth-generation TPUs.

Reliability for AI supercomputers

TPU superpods consist of thousands of chips arranged into cubes (64 TPUs), with high-speed Inter-Chip Interconnect (ICI) links connecting every chip within a cube and a dynamically configurable Optical Circuit Switch (OCS) network connecting all cubes to form a superpod.

For system-wide training progress, we must maximize the number of fully healthy cubes within a superpod. Because the performance of AI models relies on high-bandwidth, low-latency communication, every chip and ICI link within a cube must be operational for that unit to contribute to the training progress. Driven by these architectural realities, our cluster-level framework helps define how the industry can achieve reliability in the AI era, moving from instance-level reliability to availability of scale.

Deep dive: The mathematics of availability at scale

Instance-level reliability models are often deterministic, but industrial-scale AI deployments require a probabilistic approach over thousands of chips. In a traditional setup, you might track the Mean Time Between Failures (MTBF) of a single chip. However, at the scale of frontier AI, the cluster-level MTBF drops sharply as the number of components grows.

To visualize how quickly scaling can erode confidence, we can look at simple bounds like Markov’s inequality.

If we define X as the number of failed cubes, Markov’s inequality reminds us that as the expected number of failures E[X] increases with cluster size, the probability of staying below a strict failure threshold becomes increasingly difficult to guarantee without systemic architectural changes.

While Markov’s inequality provides a helpful rule of thumb for the risks at scale, we model the availability of scale using a binomial distribution of aggregate cluster health. For a superpod composed of n independent units (cubes), we define the probability of having at least k fully operational and interconnected cubes as the cumulative distribution of the success of n independent trials. To ensure a 95% confidence interval for training productivity, we solve for k where:

Where n represents the total cubes in a superpod and p represents the aggregate cube-level availability.

This model replaces the instance-level model with a topology-aware framework that mirrors actual performance requirements of large-scale training, ensuring that the larger block of compute is healthy and connected and can drive continuous training progress.

Scale of modern AI hardware

To demonstrate this new reliability model, we used Ironwood, Google’s generally available, seventh-generation TPU, and the custom silicon behind advanced models like Gemini and Nano Banana.

Pictured: Part of an Ironwood superpod, directly connecting 9,216 Ironwood TPUs in a single domain.

An Ironwood superpod is a dense, high-performance fabric consisting of 9,216 chips integrated into a single compute domain. It’s organized into 144 cubes, where each cube contains 64 chips. Within these cubes, ICI links create an extremely dense, all-to-all network fabric that provides massive bandwidth and low-latency connectivity for distributed operations within the cube. To form a superpod, 144 cubes are connected using OCS. For large jobs, capacity can be provisioned by interconnecting multiple cubes within a pod into one super-slice, or connecting multiple slices to form a multi-slice cluster. Cubes across multiple superpods can be connected over the datacenter network to run even larger workloads.

Using this model, we determine that the topological availability for an Ironwood superpod is 130 out of 144 cubes available for 95% of the month. This translates to a large compute block of 8,320 chips that are fully operational and interconnected via ICI and OCS, establishing a reliability model specifically optimized for hero jobs (the massive training runs of frontier AI).

The relationship between cluster size and its statistical availability is non-linear. By adjusting the required confidence level, we can identify the slice size that can be supported with statistical certainty. For researchers, this mapping provides a capacity availability curve. An organization with a workload that requires 99% availability for a mission-critical run can optimize their slice size to 125 cubes, while those pushing utmost scale can utilize 130 cubes at the 95% confidence interval.

Capacity availability curve for an Ironwood superpod (144 cubes)

This new reliability model maximizes the utility of the entire superpod through:

Full access: This model does not constrain capacity utilization; it focuses on the availability of fully healthy cubes. While a single chip or ICI failure results in the entire cube being classified as unhealthy, customers continue to have access to the remaining capacity within the cube. This makes most of the Ironwood superpod available for use while also optimizing the compute footprint for high-stakes, large-scale training.
Optimized resource usage: While the 130-cube model focuses primarily on large-scale training runs, the full superpod remains available for a heterogeneous mix of workloads. This allows researchers to utilize the remaining cubes for research experiments, inference, and dev/test workloads, maximizing the utility of the superpod without compromising the reliability of the main training run.

Our customers are using Ironwood at scale today and this model has empowered them to train their most demanding hero jobs.

Enhancing ML productivity

The goodput metric is the primary measure of ML productivity. Our new reliability standard provides the deterministic foundation for goodput and is engineered to maximize this metric for demanding hero jobs, so that the massive scale infrastructure required for frontier research is ready to perform as a single entity.

This model achieves high scheduling goodput, one of the three goodput metrics, by making the full set of resources available for massive-scale training runs. Combined with the software stack, this infrastructure-level availability helps deliver the high overall goodput. We achieve this through a three-layer reliability model:

Infrastructure: TPU superpods provide the capacity footprint to ensure the necessary scale is physically available and connected.
Frameworks: JAX and Pathways provide resilience, reconfiguring or hot-swapping around failed nodes to maintain forward progress without requiring a full restart.
Application: Fault-tolerance mechanisms like auto-checkpointing and multi-tier checkpointing preserve training state, so that lost progress is minimized in case of a failure.

Enabling the next generation of AI breakthroughs

The cluster-level reliability model marks the beginning of a new standard for the AI era, where an AI supercomputer is a dependable, industrial-scale engine for innovation. By aligning our reliability posture with the demands of frontier models, we’re enabling the next generation of AI breakthroughs to be faster, more reliable, and more predictable. Click here to learn more and get started with TPUs.

Gemini 3.1 Flash-Lite is now generally available on Gemini Enterprise Agent Platform

Thu, 07 May 2026 18:00:00 +0000

Today, we’re thrilled to announce that Gemini 3.1 Flash-Lite, our fastest and most cost-efficient Gemini 3 series model yet, is now generally available.

Designed for ultra-low latency, high-volume tasks, and unmatched cost-efficiency, Flash-Lite is already transforming how applications are built at scale. Fast, iterative, and scalable, it joins our comprehensive suite of Pro and Flash models to provide the exact combination of intelligence, speed, and cost required for the most demanding production deployments.

Developers and enterprises have noted that the model provides the precision required for agentic tasks like tool calling and orchestration, coupled with the cost-efficiency needed to run automated pipelines at scale.

Here’s a look at how some of them have been driving value.

Software development and engineering

Engineering teams require models that can keep pace with real-time coding environments. With the GA of Gemini 3.1 Flash-Lite, developers are unlocking the instant responsiveness necessary for complex code completion, seamless UX design, and agentic developer tools.

“Integrating Gemini 3.1 Flash-Lite has transformed the responsiveness of our IDE AI assistant & Junie agent. The balance of high intelligence and minimal latency makes it the perfect model for real-time developer support." — Vladislav Tankov, Director of AI at JetBrains

Customer experience and high-volume service

For enterprise customer service operations, handling massive volumes of interactions requires models that can scale affordably without sacrificing reasoning capabilities.

Gladly runs customer service for some of the most demanding retail brands in the world. The core of its text-channel AI agent runs on Flash-Lite. By handling millions of customer-facing calls each week across channels like SMS, WhatsApp, and Instagram, they achieved roughly 60% lower costs than comparable thinking-tier models on the same token mix.

The model powers every step of the agent lifecycle — from selecting tools and classifying playbooks to deciding when to escalate to a human — all while maintaining a p95 latency around 1.8s seconds for fully reply generation and sub-second p95 for classifiers and tool calls, alongside a ~99.6% success rate under heavy concurrent load.

Creative pipelines and gaming

In the fast-paced creative and gaming industries, multimodal capabilities and ultra-low latency are essential for keeping users engaged and content pipelines flowing. Flash-Lite is empowering platforms to process rich media and generate hyper-personalized environments.

Astrocade lets anyone create games by describing what they want in natural language. They integrated Flash-Lite to serve a rapidly growing global user base.

For every incoming game request, it performs a multimodal safety check — analyzing both text and images — before the building agents even start their work. It further supports their global community through inline comment translation, allowing players in different countries to “riff” on the same game. And as part of their asset generation pipeline, it helps refine the final prompts to ensure consistently high-quality thumbnails.

The creative platform krea.ai has also seen positive results by using Flash-Lite as a prompt enhancer in their Nodes tool. By taking a user’s rough idea and expanding it into a full image generation prompt pipeline, the model provides a level of detail that is “weirdly creative” for its price point.

These outputs move the needle on image production, providing a level of reliability and scale that was previously cost-prohibitive for sophisticated prompt engineering.

Financial services and data operations

In the world of finance and enterprise product development, efficiency is just as critical as accuracy. Gemini 3.1 Flash-Lite gives financial analysts and product managers the ideal balance of intelligence, low latency, and cost-effectiveness to run modeling and latency-sensitive applications.

OffDeal uses Flash-Lite to power “Archie,” an AI agent that investment bankers use for real-time research, data lookups, and task execution during Zoom calls. In these scenarios, bankers often need to surface financials mid-conversation. OffDeal found that Flash-Lite was the only model capable of meeting the response times needed for genuinely instant answers without forcing a tradeoff on quality.

Beyond live calls, they also use Flash-Lite as a triage layer for inbound and outbound email traffic. By answering structured questions about messages in parallel, such as whether an email is an automated response or in relation to an active deal, Flash-Lite determines which downstream AI agents get invoked and with what context.

For high-volume, latency-sensitive workflows on the financial operations platform Ramp, Flash-Lite has become a key component:

“Gemini is a core part of the model stack we use across applications at Ramp. As indicated in our benchmarks, we see Gemini lead the pareto fronts in terms of costs, latency and intelligence—providing a great tradeoff between the three and making it well-suited for latency sensitive applications. Gemini 3.1 Flash-Lite has been especially valuable, powering many of our highest-volume, latency-sensitive features without compromising on quality.” – Anton Biryukov, Applied AI Engineer, Ramp

Market intelligence platform AlphaSense integrates Flash-Lite to deliver data insights:

"Gemini 3.1 Flash-Lite provides great balance of speed, cost and performance, allowing AlphaSense to scale our advanced data processing and deliver high-quality intelligence across every layer of our data stack”– Chris Ackerson, Senior Vice President of Product, AlphaSense

Get started

Read the docs for Gemini 3.1 Flash-Lite and learn about our latest pricing structure. Learn more about the Gemini Enterprise Agent Platform, the new standard for enterprise agent development.

How BASF manages thousands of supply chain decisions with AlphaEvolve’s agentic algorithms

Thu, 07 May 2026 16:00:00 +0000

The agricultural and crop protection supply chain is one of the most intricate networks in the world. It takes up to two years to turn active ingredients into the final products farmers need, and a single change in weather or regulations can disrupt everything. Planners at BASF Agricultural Solutions navigate this reality daily across 180 production sites. To understand how local decisions ripple across their entire global network, BASF turned to AlphaEvolve on Google Cloud to build a digital twin of their supply chain.

Planning across a two-year lead time

BASF Agricultural Solutions manages a network with over 5,000 distinct value chains. Creating a single end product requires a bill of materials that can be over 30 levels deep, moving across different production sites and regions.

Currently, human planners make thousands of local decisions every day. They decide what to produce, when to produce it, and how much safety stock to hold. Because the network is so large, a planner can’t easily see how a localized decision affects the rest of the global supply chain.

This scale can lead to additional working capital and inventory and or cause production imbalances. Traditional mathematical models struggle to capture the dynamic reality of the network that planners navigate based on years of experience.

Building a foundation for decision support

AlphaEvolve is an evolutionary coding agent that generates and refines algorithms autonomously. In collaboration with Google Cloud and prognostica GmbH, BASF’s objective was not to replace human decision-making, but to establish a new model for decision support that helps planners handle the real-world complexity of the production network.

The team gave AlphaEvolve a foundational "seed" program. This initial code established a standard planning logic that translated demand forecasts into production schedules, serving as a functional baseline before introducing dynamic, network-wide coordination. From there, they fed the model three years of historical data, including inventory levels, market demand, and actual production outputs. AlphaEvolve then generated variations of the code, mutating the logic to see if it could simulate a supply chain that matched the real-world historical data.

Measuring what good looks like in initial tests

For AlphaEvolve to improve, it needed a specific goal. The evaluation function scored every new piece of generated code on one primary metric: how closely the simulated inventory levels and production decisions matched the actual historical reality recorded by BASF.

The latest AlphaEvolve runs delivered more than 80% relative improvement in accuracy compared to the initial seed model. With further adjustments, the team expects to push performance even higher — bringing the model to a level of accuracy not achieved with other approaches and making it actionable for operational use.

The results

The evolved planning logic delivered immediate, measurable improvements over the initial seed model. The final algorithm successfully mirrored the actual historical performance of the supply chain, significantly reducing the error rate compared to the initial seed.

“We had several attempts to build a digital twin for our complex supply network using deterministic models, and all of them failed,” said Dr. Goetz Krabbe, vice president for global supply chain at BASF. “By using AlphaEvolve, we cannot only map the complex network based on system data, but at the same time understand and copy the human decisions that drive our daily operations. This gives us a highly accurate and easy to maintain data driven digital twin of the entire network. Using it we can optimize our inventory levels and respond to market volatility with confidence while avoiding stockouts."

What the evolved algorithm actually does

By running thousands of experiments, AlphaEvolve developed a clear, human-readable algorithm that explains how the BASF network truly operates. It automatically discovered factually correct, domain-specific supply chain rules that explain the observed production outputs and inventory levels for the tested product value chain:

Production consolidation: The algorithm learned to group production amounts together, accurately mapping how planners optimize plant time.
Dynamic safety stocks: It introduced safety stock parameters to handle volatile and seasonal demand patterns, helping to strictly manage capital costs while preventing out-of-stock situations.
Network-wide coordination: The model successfully mapped the dependencies between different production tiers, providing a clear foundation for optimizing asset utilization globally.

What's next

The initial simulations showed that evolutionary AI can accurately model large-scale, dynamic supply chains. BASF’s objective is to create a digital twin of their entire global production network as a new foundation for simulation, decision support, scenario forecasting and optimization. This will allow the team to continuously simulate operations, identify hidden bottlenecks before they affect throughput, and optimize asset utilization across all global facilities.

_{This project was a collaboration between the BASF SE team including: Benjamin Priese, Michael Arlt, Debora Morgenstern and Tobias Hausen as well as Manuel Doerr and Thomas Christ from Prognostica GmbH Würzburg, and the AI for Science team at Google Cloud including (but not limited to): Kartik Sanu, Laurynas Tamulevičius, Nicolas Stroppa, Chris Page, Srikanth Soma, John Semerdjian, Skandar Hannachi, Vishal Agarwal and Anant Nawalgaria as well as Christoph Tittelbach from the Google account team and partners at Google DeepMind}

Fitting the future: How Breuninger boosted sales with its "be your own model" AI

Wed, 06 May 2026 16:00:00 +0000

“How will this look on me?” It’s the question every online fashion shopper asks, and one that most retailers still can’t answer well.

Breuninger, a fashion and lifestyle company based in Germany, thought emerging generative media models could be a good fit for this fashion conundrum. Working with Google Cloud, they built a virtual try-on experience that lets shoppers see high-end fashion on their own bodies using a simple selfie.

From trusted tester to live product

The project began when the Google Cloud team in Germany invited Breuninger to join the Trusted Tester Program for the Virtual Try-On (VTO) API. Breuninger’s data team in Germany worked directly with Google’s engineers in California, testing and refining the technology in three stages:

Catalog enrichment: The team first explored the VTO API to dress professional models in different outfits. This helped Breuninger to cover a greater variety in user tests without having to plan new photoshoots.
Body type selection: They then added a feature that let users choose from different body types to see how clothes would drape on a silhouette similar to their own.
The 'Be your own model' breakthrough: User feedback showed that customers did not just want to see a model; they wanted to see themselves.

The product owner at Breuninger noted that this close collaboration allowed the team to share user feedback with developers in real time. This speed helped them move from using pre-selected models to a user-first, selfie-based approach.

Three levels of virtual try-on

The project revealed three levels at which retails can adopt VTO, depending on how much personalization they want to offer:

Approach	Interaction	Use case
Level 1: Catalog enrichment	Offline batch processing	Dress standard models in new collections at scale to update product pages without manual shoots.
Level 2: Body type selection	Online on-request	Offer predefined models for users to choose from, similar to the virtual try-on feature on Google Shopping.
Level 3: 'Be your own model'	Online personalized	The most personal experience where users upload a selfie to see themselves in specific items or full outfits.

Building for scale with Flutter

Scaling a personalized experience required more than just an AI model. Selfies come in wildly different lighting and quality, so the team built preprocessing tools to make sure the final images met Breuninger’s brand standards. This project also accelerated Breuninger’s move to a Flutter-based platform. The VTO feature was the first module built by a self-sufficient product team using this new structure, which helped the team move from a vision to a live launch in only three months.

Real results during the holiday season

During a six-week A/B test over Black Week and the holiday season, the team found that shoppers who used the virtual try-on converted at a higher rate and generated a stronger contribution margin than those who didn't. Customer surveys reinforced the numbers: shoppers responded well to the high image quality and the personalized experience. Perhaps most telling, the team found that VTO became a tool for building style confidence — helping customers feel sure about a purchase before they made it.

What’s next

The pilot’s success has set up a broader rollout and international expansion, with physical fit and sizing support on the roadmap. Breuninger continues to refine the experience based on how customers actually use it in everyday shopping — the same user-first approach that shaped the project from the start.

To explore how generative AI can help your business create similar experiences, visit Google Cloud's Virtual Try-On solution. You can also try the feature yourself in the Breuninger app.

_{This work wouldn’t have been possible without the contributions from peers at both Breuninger, and Google Cloud. Thanks to Markus Peetz, Jorina Hilser, Martin Csengeri, Jay Deutinger, Sofia Widmayer, David Schowalter, Tobias Götze, Eric Karge, Abdul Mateen, Besnik Brahimi, Oliver Fesseler, and Lisa Beutner from Breuninger, and Khanh LeViet, Jorj Ismailyan, and Matt Chaban from Google Cloud.}

Pioneering AI-assisted code migration: How Google achieved 6x faster migration from TensorFlow to JAX

Wed, 06 May 2026 16:00:00 +0000

AI coding agents are rapidly becoming ubiquitous across the software industry, fundamentally changing how developers write, test, and debug daily code. While these tools excel at localized, self-contained tasks, applying them to massive, systemic codebase migrations requires an entirely new approach.

Google is already addressing this challenge by incorporating AI into many migration workflows: x86 to ARM (enabling workloads on Google Axion processors); int32 to int64 identifiers (to avoid running out of ids); JUnit3 to JUnit4 (for testing); and Joda-Time to java.time (a modern time library). However, AI model migration represents a whole new level of complexity that requires even more advanced methods for AI-assisted migration.

Translating a production-grade machine learning model from one framework to another, for example, from TensorFlow (TF) to JAX, is not a simple syntax update. It is a long-horizon task that requires untangling thousands of lines of code, managing complex states across multiple files, and preserving precise mathematical equivalence. Generic, single-agent coding assistants typically struggle under this weight — they frequently lose context over long workflows, hallucinate APIs, or fail to produce buildable code across an entire repository.

Google’s AI and Infrastructure team has pioneered a new approach to this industry-wide problem. The result is 6x faster model migration, a milestone Sundar highlighted in the recent Google Cloud Next keynote. In this post, we share how we deployed specialized, multi-agent AI systems to migrate some of Google’s largest-scale production models from TF to JAX.

Accelerating the transition from TF to JAX

For many teams at Google — and across the industry — the future of scalable machine learning is being built on JAX. Designed around a functional, stateless paradigm, JAX is heavily optimized for modern Tensor Processing Unit (TPU) infrastructure and XLA compilation, making it the bedrock of the modern AI stack.

Evolving to this future presents a monumental challenge. Thousands of production models are built on TensorFlow, a framework characterized by object-oriented, stateful layer initialization and static execution graphs. Manually migrating these models to JAX requires a fundamental rethinking of how layers interact, and how state is explicitly managed. Across large organizations, this type of migration alone represents hundreds (if not thousands) of software engineering (SWE) years — time better spent on researching new architectures and driving product innovation.

Overcoming this challenge with AI started as an ambitious experiment within Google’s AI and Infrastructure team, but has evolved into a repeatable blueprint for addressing complex engineering problems across the company.

Moving beyond single-agent coding

Our early experiments with agentic code translation showed promise for simple models. However, when faced with the realities of a Google-scale migration — complex, production-grade models spanning multiple files and thousands of lines of code — generic, single-agent setups struggled. They could not balance high-level structural rules with low-level execution details, resulting in a variety of failures, such as overwriting critical files or skipping necessary functionality. To overcome these common challenges inherent to enterprise migrations, we developed a highly specialized multi-agent architecture that consists of:

The Planner agent: Using deterministic, compiler-based static analysis, the Planner maps out the codebase's entire dependency tree. It then works alongside other agents to break the migration down into a discrete, step-by-step plan, helping ensure the migration happens logically from the "leaf nodes" (layers without unmigrated dependencies) upward.
The Orchestrator agent: This agent acts as the project manager. It dynamically groups plan steps into manageable chunks to keep the context window focused, injects the necessary domain knowledge, and handles failure recovery if a step doesn't build.
The Coder agent: Built as a reasoning and acting agent, the Coder is the workhorse. Integrated directly into our internal IDE tools, it has the ability to read files, write code, run builds, and execute unit tests. Crucially, it operates in a "test-and-fix" loop, self-correcting until it produces a compilable, verifiable component in the target language.

Figure: Multi-agent AI system for complex code migrations. Process diagram describing the multi-agent system used to migrate legacy model code to JAX. Image generated with Gemini Nano Banana 2.

Scalable validation and dynamic Playbooks

Generative AI models are only as good as the context they are provided. Because source and target architectures rarely map 1-to-1, we engineered a scalable, hierarchical system of Playbooks.

These Playbooks range from general repository instructions to highly specific "golden examples" distilled from successful manual migrations. By feeding the Orchestrator a client-specific Playbook (for instance, one tailored to YouTube's unique ranking model infrastructure), the system avoids generic hallucinations and strictly adheres to internal coding standards. This Playbook architecture is framework-agnostic, meaning it can be adapted to guide migrations between any two programming languages or frameworks.

Furthermore, we instituted rigorous quality metrics to ensure the generated code is actually production-ready:

Quantitative verification: For each unit of code, we verify correctness mathematically. In the case of the TF-to-JAX migration, the system utilizes algorithmic gradient ascent to find the maximum error between the original TF layer and the new JAX layer, mathematically verifying functional equivalence.
Qualitative evaluation: We also evaluate the migrated code against a set of qualitative standards. In the case of the TF-to-JAX migration, we deploy a blind-audit LLM Judge that scores the migrated code against a framework-agnostic architectural checklist, so that critical, domain-specific logic is completely captured.

Redefining migration velocity

By deploying this multi-agent system, we dramatically alter the economics of software migration.

In our evaluations on real-world, highly complex YouTube models (featuring thousands of lines of code, hundreds of layers, and deep metric dependencies), the multi-agent system achieved a 6.4x to 8x speedup over performing the migration manually. What traditionally took several SWE-months can now be reduced to only a few weeks of AI-assisted code generation, followed by expert human review.

The system effectively handles the boilerplate, identifies target idioms, maps the dependencies, and generates the unit tests, allowing engineers to act as reviewers and architects rather than manual translators.

Looking ahead into the AI-assisted era

AI is transforming the pace of technological innovation. Without using AI to accelerate our ability to conduct large-scale migrations, it will become increasingly difficult for organizations to adopt the latest breakthroughs and maintain the security, reliability, and performance of their systems.

Our work migrating machine learning implementations from one ML framework to another demonstrates that by combining deterministic static analysis, strict testing loops, and specialized multi-agent architectures, we can safely automate some of the most complex software engineering challenges in the industry. A detailed description of the process is published in our technical paper.

_{This work is the result of collaboration across Google. We thank key contributors: Stoyan Nikolov, Niyati Parameswaran, Bernhard Konrad, Moritz Gronbach, Niket Kumar, Ann Yan, Varun Singh, Yaning Liang, Antoine Baudoux, Xevi Miró Bruix, Daniele Codecasa, Madhura Dudhgaonkar, Elian Dumitru, Alex Ivanov, Christopher Milne-O’Grady, Ahmed Omran, Ivan Petrychenko, Assaf Raman, Stefan Schnabl, Yurun Shen, Maxim Tabachnyk, Niranjan Tulpule, Amin Vahdat, and Jeff Zhou.}

The Blueprint: Translating stream-of-conscious speech into responsive, actionable task lists

Wed, 06 May 2026 16:00:00 +0000

Welcome to The Blueprint, a new 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.

Founded in 2007, Doist is a pioneer in async and remote-first work on a mission to simplify life’s complexities through apps like Todoist for task management and Twist for team communication.

The challenge:

We launched Ramble to take our popular Todoist application to the next level by capturing non-stop, stream-of-consciousness talking. Our inspiration was that scene from The Devil Wears Prada where Miranda Priestly rapid-fires a dozen tasks at her assistant. We asked: What if anyone could capture tasks that way? No typing, no careful formatting. Just talk and let Todoist do the organizing. That use case became our north star.

At the outset, we identified four big technical hurdles:

We needed fast and accurate real-time communication with tool-calling capabilities.
Multilingual support at scale but with great support for slang, accents, and more.
As traditional assertion-based testing would not work for our platform, we would have to find a way to achieve non-deterministic output testing and semantic validation.
Reliable, flawless handling of audio across browsers.

The solution:

We built Ramble using Gemini Enterprise Agent Platform and its previous iteration, Vertex AI; specifically, we’re using Agent Platform to access the Gemini Flash models. We chose these over other options primarily due to the quality of Google's state-of-the-art models and its clear terms and assurances about preserving privacy.

Gemini’s Live API (accessed via Agent Platform) powers Ramble’s core real-time interactions and key capabilities, including native audio streaming, proactive tool calling, session resumption, and multilingual understanding.

Ramble sends the raw PCM audio directly to the model without pre-transcription. Gemini handles language detection, speech recognition, and semantic understanding in a single pass, reducing latency. It then invokes our purposefully designed tools (addTask, editTask, deleteTask, etc) autonomously as the user speaks, without waiting for explicit commands.

The APIs in Agent Platform provide resumption tokens that let users pause and continue sessions, which is essential for mobile users who might switch apps or lose connectivity.

The end result is a clear, concise list of the tasks, regardless of how many, how inconsistently, or how confusingly they may have been rambled by the user.

The architecture:

The outcome:

Ramble has come to rely on the quality of Google’s AI models, particularly the reasoning and near-instant audio-processing capabilities of Gemini Flash. Other platforms and models offer similar capabilities, and we did bake in support for them, but none hit our internal quality bar as consistently as Gemini. When it came to a user's unstructured “rambling” and the need to fill in gaps, Gemini turned out to be the most intelligent of all the models we explored. The result was the clearest and most consistent breakdown of tasks, which was the exact magical user experience we wanted to create.

After an early rate-limit incident caused by unexpectedly high usage during alpha testing, we developed a deeper, more proactive partnership with Google, ensuring long-term sustainability and the support necessary for our high API usage. Since then, it's been easy for us to connect directly with Google Cloud staff, including engineers, when issues arise.

Here at Doist, Ramble took off both in a qualitative and quantitative sense. It’s become a hallmark experience that incentivizes us to explore tasteful applications of AI that can enhance our existing product experience, both in the B2C space as well as B2B. Beyond task creation, we’re considering several opportunities across the productivity journey, from capture to planning and even automation.

The details:

We structured our back-end to enable future voice-powered features. The architecture includes a provider-agnostic streaming layer; a dictation module for one-way audio; Ramble (our “brain dump” module); and a conversation module to support streaming bi-directional audio and future conversational features.

This layered design means we can ship new voice features with minimal additional infrastructure work. It also enables provider flexibility; although we’re using Gemini Enterprise Agent Platform in production, our abstraction layer also easily supports other solutions.

In addition to helping us tackle three of our four key technical challenges, Agent Platform delivered some nice surprises. First, session resumption was easier than we expected. We initially thought maintaining conversation state across reconnections would require complex server-side session management. But once we understood Agent Platform’s resumption token approach (the token is provided by the API and changes with each context update), implementation was straightforward across all platforms.

Second, context injection worked on the first try. We spent considerable time designing how to provide user context (projects, labels, preferences) to the model. We explored complex retrieval strategies and dynamic context windows. In the end, the simple "v1" approach—just passing most of the user's metadata in the system prompt—worked remarkably well.

For testing, we combined structural validation (task count, priority levels, date presence, etc) with semantic validation (did the model understand the user's intent?) following the LLM-as-judge approach. A second Gemini model evaluates whether the output semantically matches the expected outcome. Native speakers from our global team recorded real-world scenarios in their languages and local accents (15+ language variations and over 100 recordings total), with each scenario having expected semantic outcomes (e.g., "should create 3 tasks: one about calling family, one about shopping, one about exercise on Saturday at 11 AM").

We then created a defined pass-rate threshold for the test suite overall, while also monitoring per-language performance to catch regressions. This approach lets us evaluate new model versions systematically, understanding not just overall performance but also which specific languages might see improved or degraded experiences, and make data-informed decisions.

Ultimately, Ramble is a resounding success in helping our users handle the chaos of day-to-day life. It joins the ranks of Todoist’s Quick Add — our existing natural-language task input — in providing yet another way to capture tasks that is the best in its category.

Five must-have guides to move agents into production with Gemini Enterprise Agent Platform

Tue, 05 May 2026 16:00:00 +0000

Building AI agents that work well in a demo is one thing, but running them in production requires serious infrastructure.

At Google Cloud Next '26, we introduced Gemini Enterprise Agent Platform to help developers build, deploy, scale, govern, and optimize autonomous AI agents. From managing long-running state and enforcing security with the Agent Governance Stack, to orchestrating complex workflows using Agent Development Kit, these tools help you treat your agent fleet with the same rigor as your engineering organization. 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.

1. Agent design patterns for long-running AI agents

Developers spend weeks perfecting prompt engineering, tool calling, and response latency. But none of that matters when your agent loses its reasoning chain over a five-day task. At Next 26, we announced that Agent Runtime now supports long-running agents that maintain state for up to seven days.

In this article, we’ll share five essential agent design patterns for building long-running agents with Agent Runtime. You’ll learn how to implement checkpoint-and-resume mechanisms to recover from failures without starting over. We also cover how to build delegated approval workflows where the agent pauses for human review while consuming zero compute resources.

Read the full guide on long-running agents here.

2. The agent governance stack

A misconfigured SaaS tool leaks data passively, but a misconfigured agent takes bad actions actively. The pattern we saw with shadow IT in 2015 is repeating itself with AI agents.

To manage this risk, we explain why you must treat your agent fleet with the same rigor as your engineering organization. We outline a five-layer governance stack designed to provide your r security team with precise visibility and control. The foundation begins with Agent Identity, assigning every agent a unique cryptographic badge to isolate access. From there, we explore how to use Agent Registry for centralized tool governance and Agent Gateway to enforce natural language security policies across your fleet. The stack concludes with behavioral anomaly detection and a unified security dashboard to monitor your overall risk.

Read the full guide on the agent governance stack here.

3. Must-have multi-agent orchestration patterns in ADK

Building a single AI skill is relatively straightforward, but orchestrating multiple skills across different agents is notoriously difficult. With the new updates to Agent Development Kit (ADK), we introduced graph-based workflows, collaborative agents, and a formalized skills framework to solve these orchestration failures.

Our third guide details five multi-agent orchestration patterns you can use to build reliable systems. You will find code examples for building hybrid graphs that combine hard-coded business rules with flexible AI reasoning. We also show how to use the coordinator-specialist pattern to avoid building monolithic, unpredictable agents. The guide concludes with deep dives into skill composition, cross-language pipelines, and secure sandboxed executors for running arbitrary code.

Read the full guide on ADK multi-agent patterns here.

4. Deep dive: How A2A and MCP work together

Organizations will rarely build every AI agent they need entirely from scratch. The real value comes when agents built by different teams, in different languages, and across different organizations can securely discover and collaborate with each other.

In our final guide, we explore five integration patterns using the Agent-to-Agent (A2A) and Model Context Protocol (MCP) standards. You will see how Agent Cards allow agents to publish their capabilities so coordinator agents can find them through the Agent Registry. We also show how MCP acts as a universal tool bridge to connect your agents to databases and enterprise systems without custom integration code. The article finishes with strategies for cross-organization federation that involves agents from different organizations collaborating on shared tasks using the Agent Gallery in Gemini Enterprise and building ambient event meshes for agents that react to events continuously in the background, without waiting for user requests.

Read the full guide on agent interoperability here.

5. Atomic agent blueprints on Google Cloud’s Agent Garden

Building multi-agent systems from scratch presents complex design challenges, including finding the optimized design pattern for your use-case, orchestration failures and evaluation loops. You can spend weeks reinventing the wheel, trying to get your agents to be ready for production - or you can start with architectures that already work, with our new Atomic Agents in Agent Garden.

Read the full guide to learn about pre-built Agent Blueprints in Agent Garden

Watch the complete Agent Platform explainer

To see these architectural patterns in practice, watch this technical walkthrough of the Gemini Enterprise Agent Platform. This deep dive covers the complete agent lifecycle, showing you exactly how to move from initial code to a secure, scalable AI Agents in production.

Dive into the code with Agent Platform samples on GitHub

Access our curated repository of code samples and tutorials for the Gemini Enterprise Agent Platform. This GitHub repository provides practical examples for the entire agent lifecycle, giving you the exact code needed to build, scale, govern, and optimize your autonomous fleets.

Get started with Gemini Enterprise Agent Platform

Moving agents into production requires both robust infrastructure and the flexibility to choose the right reasoning engine for the task. The Gemini Enterprise Agent Platform bridges this gap, allowing you to build, govern, and scale autonomous workflows with complete enterprise control.

Through first-class integration with Model Garden, your agent fleet has direct access to more than 200 leading models. You can route tasks to the best available option, whether that is a first-party model like Gemini 3.1 Pro or Lyria 3, an open model like Gemma 4, or third-party models like Anthropic’s Claude, Opus, Sonnet or Haiku.

Visit Agent Platform in the Google Cloud console to explore new features and start building today.

Introducing Agent Gateway ISV ecosystem for security and governance

Tue, 05 May 2026 16:00:00 +0000

Managing agents and their actions can quickly grow in complexity and introduce security risks unique to AI. To address these challenges, at Google Cloud Next we announced Agent Gateway to provide simple, secure, and governed connectivity across all user-to-agent, agent-to-agent, and agent-to-tools interactions.

As part of Gemini Enterprise Agent Platform, Agent Gateway provides a programmable data plane for your AI agents. It connects easily with a wide array of security providers, giving your team the flexibility to inject custom logic and third-party security controls directly into the request path.

To support the agentic enterprise in today’s multicloud and multi-AI world, we’re partnering with leading identity and AI security providers to integrate with Agent Gateway and help ensure that your security posture remains as flexible as the agents you’re building.

Agent Gateway partner ecosystem for agent security and governance.

Broadcom: Agentic AI introduces high-speed, autonomous data exchanges across LLMs, tools, and other agents, dramatically expanding the risk of data exfiltration through new, unmonitored leakage points. To counter this, Symantec and Google Cloud are partnering to integrate Symantec Data Loss Prevention (DLP) scanning as a service extension for the Agent Gateway, which serves as the network-level enforcement point for all agent traffic. This integration enables real-time inspection and enforcement of existing DLP policies across agent communications — including LLM inference requests and MCP tool calls — without requiring any changes to application code.
Check Point: Securing your AI transformation across both employee adoption and runtime innovation, Check Point’s AI Defense Plane can discover and govern sanctioned and unsanctioned, shadow AI usage. AI Defense Plane’s runtime protections integrate with Agent Gateway to provide low-latency inspection of prompts, responses, and tool interactions — preventing agent manipulation, sensitive data leakage, and tool misuse, so organizations can confidently scale AI.
Cisco: Integrating Cisco AI Defense with Agent Gateway can help enforce runtime protections for every AI interaction, including those that use model context protocol (MCP). These guardrails can help mitigate threats like prompt injection and data exfiltration, and agent-specific risks like tool exploitation and misuse.
CrowdStrike: Extending the AI-native CrowdStrike Falcon platform into the Agent Platform including Agent Gateway ecosystem can help CrowdStrike deliver guardrails, visibility, and control as agentic AI systems move from experimentation into production. Integrations including CrowdStrike Falcon AI Detection and Response (AIDR) and CrowdStrike Falcon Shield can provide secure operation of agents across the ecosystem.
Exabeam: Delivering behavior‑driven security analytics at enterprise scale, Exabeam New‑Scale Analytics is purpose‑built to secure Google AI and Agent Platform environments. Exabeam can ingest and analyze telemetry from Agent Platform including Agent Gateway, applying behavioral analytics to identify anomalous and high‑risk AI agent activity. Together, Google provides the AI infrastructure and controls, and Exabeam delivers the enhanced behavioral intelligence, governance, and continuous security oversight required to operate AI agents safely at scale.
F5: F5 AI Guardrails provides runtime protection for agents against data leakage, harmful outputs, and adversarial attacks. Integrated via Agent Gateway, it enforces data security and policy controls to ensure agent interactions remain governed and compliant across all models.
Netskope: Netskope One DLP On Demand with Agent Gateway inspects data at the precise moment it moves through your AI workloads and enforces the data security policies your team has already built. By embedding DLP in their architectures, organizations can govern sensitive data generated and routed by AI agents without creating new configurations, ensuring data security evolves alongside cloud and AI innovations.
Okta: Okta for AI Agents provides centralized identity governance and access control for Agent Gateway. With Okta as the identity layer, Google’s policy engine can defer access decisions to Okta, enabling organizations to govern which users and agents can access specific agents and tools. Agents created in Google Cloud can also be automatically registered in Okta, keeping identity and governance policies in sync.
Palo Alto Networks: Deploying Palo Alto Networks Prisma AIRS as an AI security layer with Agent Gateway can provide the real-time security and governance necessary to oversee agentic interactions and intercept adversarial attacks on AI before they can compromise the system. This architectural integration can help ensure that as you scale your autonomous agents, every agentic action is validated against enterprise safety and security policies, providing comprehensive operational integrity without hindering the speed of innovation.
Ping Identity: Ping Identity integrates with Agent Gateway to bring runtime identity and real-time, fine-grained authorization to agent and tool traffic. The integration with Agent Gateway ensures every request is continuously verified based on user, agent, context, and policy, rather than relying on static credentials. Together, they provide centralized, consistent governance and visibility across all agent interactions without requiring changes to application code.
Saviynt: Saviynt provides identity security and governance that helps enterprises govern every identity — human, non-human, and AI — across cloud environments. Saviynt’s integration with Agent Gateway provides live identity intelligence for every AI agent access request, evaluating intent, data sensitivity, and organizational policy in real time before access is granted. This ensures AI agents remain purpose-bound and continuously governed, with high-risk actions surfaced for human oversight and a defensible audit trail for compliance.
Silverfort: Silverfort provides identity security for agentic workloads by extending its patented Runtime Access Protection (RAP) to agent platforms, automatically discovering AI agents, mapping each to its human owner, and surfacing risks such as overprivileged access and stale credentials. By integrating directly with Agent Gateway, Silverfort can authenticate and authorize every agent-to-resource request at runtime, blocking unauthorized actions before they reach downstream systems.
Thales (Imperva): Thales provides advanced web application and API security for the Agent Platform, including security for client‑to‑agent traffic leveraging Agent Gateway. Imperva for Google Cloud (IGC), currently in preview, deploys natively in Google Cloud, eliminating the need for external software-as-a-service (SaaS) integrations and avoiding traffic redirection outside of Google’s infrastructure.
Zscaler: Providing runtime protection and governance for AI apps, models, and agents, Zscaler AI Guard can help enable real-time inspection of prompts and responses to detect malicious inputs like prompt injections and prevent sensitive data leakage through advanced content moderation and data protection detectors. The Zscaler AI Guard integration with Agent Gateway can help ensure that agentic workflows remain secure, compliant, and aligned with enterprise security policies.

As enterprises build and deploy a wide range of agents and agentic use cases, Agent Gateway supports a wide variety of agentic security controls tailored to your unique operational needs. Our approach can help your business meet compliance and governance requirements, while offering the freedom to use your choice of security provider.

To learn more about how our partners can elevate your Google Cloud experience, reach out to our team for a personalized consultation and discover the power of an open, integrated approach.

What Google Cloud announced in AI this month

Thu, 30 Apr 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.

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.

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 0x7f659c43d880>), ('btn_text', 'Start building for free'), ('href', 'http://console.cloud.google.com/freetrial?redirectPath=/vertex-ai/'), ('image', None)])]>

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.

Cloud CISO Perspectives: At Next ‘26, why we’re multicloud and multi-AI

Thu, 30 Apr 2026 16:00:00 +0000

Welcome to the second Cloud CISO Perspectives for April 2026. Today, Francis deSouza, COO Google Cloud and President, Security Products, explains why Google is multicloud and multi-AI, straight from Next ‘26.

aside_block: <ListValue: [StructValue([('title', 'Get vital board insights with Google Cloud'), ('body', <wagtail.rich_text.RichText object at 0x7f659255e280>), ('btn_text', 'Visit the hub'), ('href', 'https://cloud.google.com/solutions/security/board-of-directors?utm_source=cgc-site&utm_medium=et&utm_campaign=FY26-Q2-GLOBAL-GCP39634-email-dl-dgcsm-CISOP-NL-177159&utm_content=-&utm_term=-'), ('image', <GAEImage: GCAT-replacement-logo-A>)])]>

Cybersecurity in the era of the agentic enterprise

By Francis deSouza, COO Google Cloud and President, Security Products

Francis deSouza, COO Google Cloud and President, Security Products

Last week at Google Cloud Next ‘26, we announced 220 products, and signaled a paradigm shift. We are not just moving workloads to the cloud; we are entering the era of the agentic enterprise.

The AI megatrend, coupled with an accelerating cloud adoption, is the most profound enterprise IT transformation of our lifetimes. It is igniting a new wave of innovation, and also demands a fundamental re-architecting of cybersecurity. Our vision at Google Cloud is clear: to be the most AI-native, open, and secure platform on the planet, meeting enterprises exactly where they are.

Security at machine speed: From minutes to seconds

In this new landscape, IT resilience is defined by a multi-AI and multicloud strategy. A durable AI roadmap cannot rely on a single model or a single cloud provider. For CISOs, the mission-critical frontlines have shifted to securing models, agents, and the data that fuels them.

AI isn't just a security challenge — it is also the ultimate security tool. Today, our security operations center (SOC) agents automatically triage tens of thousands of unstructured threat reports every month. The results of our AI-first cyberdefense are transformative:

90% reduction in threat mitigation time by filtering noise and extracting intelligence instantly.
30 minutes to 60 seconds: Our Triage and Investigation agent, powered by Gemini, has processed over 5 million alerts this year, turning half-hour manual tasks into one-minute automated actions.
98% accuracy: Our new dark web intelligence capability analyzes millions of daily external events to surface the threats that actually matter.

The multicloud reality is non-negotiable

Modern organizations are multicloud by default. Between hyperscalers, SaaS vendors, and legacy systems, the single cloud dream is over. Our ethos has always been open because that is the only way to protect a fragmented world.

The reality is that AI and cloud applications are built across multiple platforms and models. To protect them, we focus on making it easier and faster to mitigate risk across all major cloud environments.

By unifying security across all major cloud environments, we aren't just simplifying management — we are lowering the stakes. Our unified approach reduces the risk and cost of a breach by 70%.

The integration of Wiz into Google Cloud has further deepened this advantage. With 90% of environments now running self-hosted AI software, Wiz allows us to secure the entire AI development lifecycle across any cloud, complementing our deep expertise in threat intelligence.

The Google advantage: From lab to live on day 1

The speed of innovation in AI is relentless. Standard security industry timelines of six months to a year to incorporate the latest models into security products are not sufficient; they leave organizations two generations behind their adversaries.

Francis deSouza, COO Google Cloud and President, Security Products, explains Google Cloud's multicloud and multi-AI approach to Next '26 attendees in Las Vegas.

Google occupies a unique position in this race. We co-design the entire stack: hardware, AI, and security.

Vertical integration: We are the only security provider that integrates a new model on day 1.
Research to reality: When Google DeepMind achieves a breakthrough in the lab, we move it to your security platform faster than anyone else in the industry.

A blueprint for the agentic future

As we advocate for a multi-AI world, we are providing the tools to build it safely. Our latest whitepaper, Building Secure Multi-Agent Systems on Google Cloud, is a robust framework for this transition.

It highlights the power of our newly announced Gemini Enterprise Agent Platform, featuring:

Agent Gateway: A single governance layer for identity and access management.
Model Armor: Sophisticated prompt sanitization to prevent adversarial attacks.
Agent Identity: Ensuring that as agents move at machine speed, they do so with authenticated authority.

The announcements at Next ‘26 were more than a recap; they were a promise. We are committed to being your partner in this new era — providing the most open, productive, and secure foundation for the AI-driven future.

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

aside_block: <ListValue: [StructValue([('title', 'Tell us what you think'), ('body', <wagtail.rich_text.RichText object at 0x7f659255ee20>), ('btn_text', 'Vote now'), ('href', 'https://www.linkedin.com/feed/update/urn:li:activity:7455362783040282624'), ('image', <GAEImage: Cloud-CISO-Perspectives-logo-A>)])]>

In case you missed it

Here are the latest updates, products, services, and resources from our security teams so far this month:

Next ‘26: Redefining security for the AI era with Google Cloud and Wiz: At Google Cloud Next, we showcased how we can help you defend against threats at machine speed, protect AI and multicloud environments, and secure cloud workloads at scale. Read more.
Next ‘26: Introducing Google Cloud Fraud Defense, the next evolution of reCAPTCHA: We’ve launched Google Cloud Fraud Defense, the trust platform for the agentic web and the next evolution of reCAPTCHA. Read more.
Next ‘26: New partner-supported workflows for Google Security Operations: We’ve introduced new partners for Google Security Operations as part of the Google Cloud Security Integration Ecosystem program. Read more.
How Google Does It: An inside look at cybersecurity: Learn how Google approaches some of today's most pressing security topics, challenges and concerns, straight from Google experts. View the collection.
The current state of prompt injections on the web: Our threat intelligence teams initiated a broad sweep of the public web to monitor for known indirect prompt injection patterns. This is what we found. Read more.

Please visit the Google Cloud blog for more security stories published this month.

aside_block: <ListValue: [StructValue([('title', 'Join the Google Cloud CISO Community'), ('body', <wagtail.rich_text.RichText object at 0x7f659255e1c0>), ('btn_text', 'Learn more'), ('href', 'https://rsvp.withgoogle.com/events/google-cloud-ciso-community-interest-form-2026?utm_source=cgc-blog&utm_medium=blog&utm_campaign=FY25-Q1-global-GCP30328-physicalevent-er-dgcsm-parent-CISO-community-2025&utm_content=cisop_&utm_term=-'), ('image', <GAEImage: GCAT-replacement-logo-A>)])]>

Threat Intelligence news

Defending your enterprise when AI models can find vulnerabilities faster than ever: Now is the time to strengthen playbooks, reduce exposure, and incorporate AI into security programs. Here’s an overview of the evolving attack lifecycle, how threat actors will weaponize these capabilities, and a roadmap for modernizing enterprise defensive strategies. Read more.
German cyber criminal Überfall and shifts in Europe's data leak landscape: Germany has reclaimed its position as a primary focus for cyber extortion in Europe. While data leak site posts rose almost 50% globally in 2025, Google Threat Intelligence (GTI) data shows that the surge is hitting German infrastructure harder and faster than its regional neighbors. Read more.
How UNC6692 employed social engineering to deploy a custom malware suite: Google Threat Intelligence Group (GTIG) has identified a multistage intrusion campaign by a newly-tracked threat group, UNC6692, that used persistent social engineering, a custom modular malware suite, and deft pivoting inside the victim’s environment to achieve deep network penetration. Read more.

Please visit the Google Cloud blog for more threat intelligence stories published this month.

Now hear this: Podcasts from Google Cloud

AI, Zero Trust, and secure by design walk into a bar: Is there Zero Trust for AI? Why is secure by design picking up speed now, just as issues of machine identity come to the fore? Grant Dasher, distinguished engineer, Google, analyzes the intersection of trust, secure design, and AI with hosts Anton Chuvakin and Tim Peacock. Listen here.
From CISA to cloud: AI assurance, concentration risk, and the new regulatory frontier: Jeanette Manfra, VP, head of Risk and Compliance, Google Cloud, joins Anton and Tim to discuss the current regulatory landscape facing cloud and AI, and the ongoing tug-of-war between security and privacy at the enterprise level. Listen here.
More than just packets: Is NDR a first-class cloud security control: Extrahop’s Raja Mukerji and Rafal Los join Anton and Tim to delve into the value proposition of network detection and response in 2026, and how it can apply to the worlds of work from home, cloud and SaaS, encryption, and high bandwidth. Listen here.
Defender’s Advantage: Takeaways from the 2026 M-Trends report: Host Luke McNamara is joined by Mandiant’s Chris Linklater to discuss the breach trends throughout 2025 and into this year. He notes key areas that organizations should focus on as we approach the mid-point of 2026. Listen here.
Cyber-Savvy Boardroom: Head in, hands out: Mark Lobel, formerly of PwC, joins hosts Alicja Cade and David Homovich to discuss why high-stakes simulations are essential to protecting corporate reputation when the regulatory clock is ticking. Listen here.

To have our Cloud CISO Perspectives post delivered twice a month to your inbox, sign up for our newsletter. We’ll be back in a few weeks with more security-related updates from Google Cloud.

The founder’s AI foundation: The top announcements for startups from Next ‘26

Wed, 29 Apr 2026 16:00:00 +0000

The momentum is undeniable: the world’s fastest-growing AI startups are building with Google Cloud. Instead of stitching together fragmented point solutions, founders are building their businesses here because we offer the entire AI stack in a single, open environment.

And, as we saw at Next ‘26 last week, we continue to advance the models, infrastructure, platforms, security, and governance that allow startups to build faster and dream bigger.

You can already see this scale and velocity in action every day. Platforms like Lovable are empowering developers to generate over 200,000 new projects daily. Thinking Machines Labs is leveraging the latest NVIDIA Blackwell chips through Google Cloud to double its training and serving speeds. When it comes to activating complex data, Parallel bypassed the multi-vendor headache, using our integrated ecosystem — bridging Gemini, BigQuery, and Spanner — to massively scale their high-accuracy search APIs.

Startups are moving at this pace because we’ve eliminated the operational plumbing. Aible is running complex enterprise agents directly where their data lives in BigQuery, while teams like Emergent AI are using Google Kubernetes Engine (GKE) to autonomously scale thousands of secure sandboxes for vibe coding natural language prompts into production-ready applications.

By putting everything in one place, we provide the ultimate foundation so you can focus 100% on shipping the AI-native products your customers are waiting for.

Let’s look at some of the biggest announcements out of Next ‘26 and what they mean for you.

1. Gemini Enterprise: comprehensive agent platform

The new Gemini Enterprise Agent Platform is how we’re moving beyond isolated AI tools to a complete lifecycle platform. It evolves the model building capabilities of Vertex AI with advanced features for:

Building: The upgraded Agent Development Kit (ADK) introduces a graph-based framework for complex multi-agent reasoning, while the low-code Agent Studio allows you to move seamlessly from prompts to deployed agents. Developers can also utilize Agent Garden to jumpstart development with pre-built agent templates.
Scaling: The re-engineered Agent Runtime delivers sub-second cold starts and supports long-running agents that maintain state for days at a time. The new Agent Memory Bank enables agents to recall high-accuracy details for personalized, long-term context.
Governing: To ensure security and compliance, the platform introduces Agent Identity for trackable auditing, Agent Gateway for unified connectivity and policy enforcement, and Agent Threat Detection to flag suspicious behavior in real time.
Optimizing: Teams can ensure quality using Agent Simulation to test against synthetic interactions, and rely on Agent Optimizer to automatically analyze real-world failures and suggest improvements.

What it means for startups: Startups no longer need to stitch together fragmented tools to build complex AI systems. The new Agent Platform provides every tool in one place — whether you are using the visual interface of Agent Studio to quickly prototype or the code-first logic of the ADK for advanced orchestration, a lean team can build and scale production-ready agents with security guardrails from day one. This approach helps eliminate technical debt and drastically accelerates time-to-market.

2. The Agentic Data Cloud: Data that actually does the work

We are moving your data from rows and columns to autonomous action. This is where you build AI agents that understand your business logic and execute the next best action—closing the gap between simply analyzing data and actually resolving real tasks on your behalf. Because agents generate orders of magnitude more workloads, our Agentic Infrastructure scales to handle bursty startup demands, improving query speeds and reducing infrastructure costs.

Query anywhere, leave data where it lives: The Cross-Cloud Lakehouse lets you use Google's advanced AI on data living in AWS or Azure. Build smarter agents without requiring migrations, experiencing vendor lock-in, or incurring additional costs to move your data.
The zero-ETL startup: Features like One-Click Reverse ETL and Lakehouse Federation collapse the walls between operational databases and analytics, allowing data to flow seamlessly without manual engineering.
Give your agents business context: Before an agent acts, it needs to understand your specific business rules. The new Knowledge Catalog turns passive data into active semantics, ensuring your agents know the exact difference between "booked revenue" and "projected revenue" without guessing or hallucinating.
Eliminate "glue code" and manual connectors: Fully managed, remote Model Context Protocol (MCP) servers are now embedded directly across the database portfolio (Firestore, Spanner, BigQuery, and AlloyDB).
Production-grade guardrails: Connecting AI to your database usually means risking dangerous SQL hallucinations. New Tools for Data Agents solve this by providing pre-built, modular building blocks that deliver near-100% text-to-SQL accuracy, letting your custom agents interact with live data safely and reliably.
Instant app generation: A new integration with Google AI Studio allows developers to use natural language for vibe coding to turn a simple text prompt into a live, deployed application connected to Firestore in seconds.

What it means for startups: Your proprietary data can be a competitive moat, but activating it usually requires a massive integration tax — breaking down cloud silos, writing brittle glue code, and managing complex ETL pipelines. Our latest data advances eliminate the friction. Startups can now leave their data right where it lives in AWS or Azure, "vibe code" an app into existence, and instantly plug agents into secure databases using standard MCPs. It turns data integration from a multi-week engineering sprint into a single afternoon's work.

3. Next-Gen Compute: Performance and Cost-Efficiency

Google Cloud is drastically expanding its AI Hypercomputer portfolio to offer increased performance-per-dollar, while introducing autonomous features that handle the heavy lifting of infrastructure management for lean teams:

TPU 8t for training: Optimized for training massive models, the TPU 8t achieves nearly 3x higher compute performance than the previous generation and delivers up to 2x better performance-per-watt.
TPU 8i for inference: Designed to make running agents more affordable, the TPU 8i delivers 80% better performance-per-dollar for inference than the prior generation.
Axion N4A for general compute: Google Cloud Axion instances deliver 100% better price-performance than comparable x86 instances, ensuring highly efficient, sustained operations.
Multi-agent networking: New C4N and M4N machine series offer a nearly 4x increase in network bandwidth, purposefully built to handle high-volume communication between AI agents.
GKE sub-second agent scaling: Google Kubernetes Engine Agent Sandbox now deploys 300 sandboxes per second, per cluster. This delivers sub-second cold starts for AI workloads, ensuring instant response times and secure multi-tenant isolation.
The self-driving cloud: Google Cloud integrated Gemini reasoning directly into its telemetry. The system now performs autonomous root-cause analysis, identifying and fixing infrastructure misconfigurations before a human even realizes there’s a problem.
NVIDIA integration: Google remains at the bleeding edge of partner hardware, becoming one of the first to offer the NVIDIA Vera Rubin NVL72 alongside Blackwell and Hopper.

What it means for startups: Managing that infrastructure can drain precious engineering cycles. Google Cloud’s expanded portfolio solves both problems. You can train models faster with the TPU 8t, scale millions of concurrent agents cost-effectively using the TPU 8i, and deliver zero-latency experiences using GKE sandboxes.

More importantly, with our autonomous root-cause analysis handling the operational plumbing, your engineers are freed up to focus on what actually matters: building your product.

4. Agentic Defense: Security as Your Enterprise Go-To-Market Engine

Google Cloud is combining its global threat intelligence with Wiz's Cloud and AI Security Platform to provide startups with a fully automated, unified security posture from code to cloud:

API economics protection: The new Google Cloud Fraud Defense acts as an intelligent bouncer, preventing malicious bots from scraping IP or running up massive compute bills on unauthorized agent interactions.
Full-stack AI protection: The Wiz AI Application Protection Platform (AI-APP) secures every layer of the AI stack and natively supports the "outer layer" of the cloud, now securing tools startups actively use like Vercel, Databricks, and Cloudflare.
Google-scale threat intelligence: New agentic SecOps tools, like the Threat Hunting Agent and dark web intelligence, automate detection rules and elevate only the threats that matter with 98% accuracy. Meanwhile, the Triage and Investigation agent uses Gemini to shrink 30-minute security investigations down to just 60 seconds.
SecOps agents: We introduced three new agents in Google Security Operations — Threat Hunting (proactive pursuit of stealthy threats), Detection Engineering (automated detection creation), and Third-Party Context (workflow data enrichment) — that empower teams to defend at the speed of AI.
Drag-and-drop security automation: Wiz Workflows introduces a new hub with a customizable drag-and-drop interface, allowing lean teams to easily orchestrate how and when these AI agents act without writing complex security scripts.

What it means for startups For B2B startups, the enterprise information-security review is where deals often go to die. By building on our unified security foundation, you can demonstrate that your application has continuous, automated red-teaming and runtime protection.

It turns the dreaded vendor risk assessment into a competitive advantage, allowing you to bypass procurement roadblocks and close enterprise deals faster without needing to hire a massive SecOps team.

5. The Go-To-Market Engine & Ecosystem Capital

Google is solving the ultimate startup hurdles — distribution and funding — by opening up its enterprise channels and deploying massive capital to partner ecosystems:

Agent Gallery in Gemini Enterprise: Startups can now monetize their customized agents directly inside the Gemini Enterprise app via Google Cloud Marketplace. This reaches millions of users right in their daily workflows and turns everyday user discovery into an automated procurement flow for IT, accelerating purchasing cycles by up to 50%.
$750-million fund: A capital injection dedicated to partner agent development and co-marketing, designed to fuel the next generation of AI builders.

What it means for startups Building the product is only half the battle; surviving enterprise procurement is the other. By integrating into the Agent Gallery, your startup can reach millions of Gemini Enterprise users directly in their daily workflow, where they can easily click to request IT procurement. Plus, their purchases can draw down on existing Google Cloud commitment, allowing you to tap into a $240 billion backlog. Paired with the $750-million partner innovation fund, Google is putting capital and distribution directly behind your growth.

Start building with the AI Agents Challenge

We’ve given you the architecture, the security, and the distribution channels — now it’s time to get hands-on. To help you accelerate your development, we are launching the Google for Startups AI Agents Challenge. Open globally to eligible startup founders and developers, this six-week competition equips your team with $500 in cloud credits and access to our latest AI tools, including Gemini Enterprise, so you can build autonomous systems and compete for a share of a $90,000 prize pool.

We’re offering separate tracks, whether you want to build a net-new agent from scratch, optimize an existing prototype for production, or prep a business-ready agent for enterprise distribution, there is a track tailored to your exact stage. Submissions are open until June 5, 2026, and will be evaluated on technical implementation, business case, innovation, and your final demo. Learn more and sign up for the challenge here.

UKG unlocks real-time workforce intelligence at scale with the Agentic Data Cloud

Wed, 29 Apr 2026 16:00:00 +0000

At UKG, we’ve spent years building and expanding our human capital management (HCM) and workforce management (WFM) solutions with new products, capabilities, and a series of acquisitions. Our cloud platform includes a suite of connected systems that support every corner of the employee experience, including scheduling and workforce operations, HR and payroll, and culture and engagement tools.

These connected tools offer customers incredible depth, but it also means our backend reflects years of evolution. We have 126 application teams, dozens of tech stacks, and more than 12,000 database instances inherited through acquisitions and product growth. And each product carries its own schema and operational footprint.

Previously, data moved through bespoke pipelines not built for real-time use. As AI advanced, expectations did too. Customers wanted instant insights across HR, time, pay, culture, and operations, and those insights increasingly needed to drive automated workflows and intelligent applications.

Internally, teams needed consistent, high-performance access to shared data to innovate faster and modernize our architecture. We needed a unified foundation for the next generation of intelligence across our suite. That’s why we built People Fabric, our new data and intelligence platform powered by AlloyDB for PostgreSQL and the just-announced Agentic Data Cloud.

Unifying the systems behind the suite

People Fabric started with a simple need: bring the full UKG suite onto one real-time foundation. Getting there started with defining a single canonical data model for the entire suite. This would serve as the shared language for people, work, pay, and culture data — consistent no matter where the information originated.

We needed an operational database that could ingest changes quickly and scale horizontally. That’s why we chose AlloyDB as the core of People Fabric. It gives us millisecond-level read-after-write behavior, high-throughput ingestion, scalable read pools, and native vector capabilities to support AI.

With the model defined and the operational store selected, the next step was building the pipeline that feeds the platform. We created a custom change data capture (CDC) framework to extract changes from our existing operational databases inherited over the years. Those changes flow through Dataflow, where they’re transformed into the canonical structure that AlloyDB for PostgreSQL expects.

Once in AlloyDB, that data becomes the real-time backbone of the platform. Applications use it for near-instant queries. AI agents rely on it for cross-domain decisions, and vector search engines use it to power natural-language and similarity-based experience layers.

For larger analytical workloads, the same data flows into BigQuery, which gives our teams and our customers the ability to perform organization-wide reporting and analysis without straining the system. Cloud SQL holds the metadata and tenancy context that govern who can see what and how different parts of the suite interact with People Fabric. From there, the system runs continuously. Data enters through streaming ingestion and gets modeled once in AlloyDB for PostgreSQL to make it available everywhere.

aside_block: <ListValue: [StructValue([('title', 'Build smarter with Google Cloud databases!'), ('body', <wagtail.rich_text.RichText object at 0x7f6590bea1f0>), ('btn_text', ''), ('href', ''), ('image', None)])]>

Bringing people intelligence to intelligent people

With the architecture in place, People Fabric gives us something we never had before: a complete and consistent view of people, work, pay, and culture data that’s updated continuously and ready for AI to use in real time.

That unified context is what powers our assistive experiences, including conversational reporting and natural-language interactions. Leaders can ask questions in plain English and get answers that reflect the full picture — not just a single system’s slice of it.

With the power of Google’s Agentic Data Cloud, our platform unifies analytical and transactional data to power real-time AI. This allows agents to reason over live workforce signals and trigger immediate actions. Because this data is governed and modeled from the start, our agents can reliably handle multi-step workflows across HR, payroll, and timekeeping.

Whether they're identifying pay discrepancies, adjusting schedules, or flagging compliance risks, they operate with the same shared semantics and security model that guides our applications. It’s the difference between AI that reacts and AI that can truly assist.

Driving impact across every layer

For engineering teams, People Fabric acts as a database-as-a-service that removes the need for each microservice to manage its own datastore or pipelines. This accelerates development and supports modernization without customer disruption.

AlloyDB for PostgreSQL delivers millisecond read-after-write behavior, zero replication lag, and near-real time ingestion latency, enabling real-time workloads with far less complexity. Migrating core person and employment data off our on-prem monolith has generated cost savings significant enough to fund half of People Fabric.

Real-time operational data now gives managers a live view of staffing, pay, and workforce activity. More than 1,000 organizations are already on the platform, with another 1,000 in progress.

As we continue expanding People Fabric, we’re laying the groundwork for deeper agentic automation, more responsive analytics, and a growing set of AI-driven capabilities — all on a trusted, scalable foundation built for what’s next.

Learn more

UKG’s success illustrates how leveraging AlloyDB for PostgreSQL and the Agentic Data Cloud allows organizations to unify operational and analytical data, creating the essential foundation for real-time, agentic AI.

Learn more about AlloyDB for PostgreSQL and get started with a free trial today!
Sign up for a strategy workshop today on how to get your data ready for the agentic era!

50+ fully managed MCP servers now available for Google Cloud services

Tue, 28 Apr 2026 17:00:00 +0000

At Google Cloud Next ‘26, we announced that more than 50 Google-managed Model Context Protocol (MCP) servers are generally available or in preview, with more on the way.

Why it matters: To move beyond experimental prototypes, AI agents must be able to access real-world data and solve complex problems autonomously.

Google-managed Managed Context Protocol (MCP) servers provide the critical connectivity to bridge AI agents with the vast Google and Google Cloud ecosystems. By hosting these servers on an enterprise-ready, standardized platform, we eliminate the need to integrate with local MCP servers and offer a unified developer experience that’s integrated across major agent runtimes and frameworks.

MCP built for the enterprise

Scaling your agent ecosystem shouldn't be a trade-off between speed and safety. You need the flexibility to grow, but you also need the guardrails to manage and govern your agents.

By pointing your AI agents toward Google-managed MCP endpoints, you’re plugging into the Google Cloud security stack, without needing to make regional configuration changes . While the platform offers deep flexibility for a range of architectures, here’s a snapshot of their capabilities that make it easier:

Strong interoperability: Through protocol translation, your agents stay compliant with the MCP specification, so that they play nice with public agents and agent frameworks like Gemini CLI, Claude, ChatGPT, VS Code, LangChain, Agent Development Kit (ADK), and CrewAI right out of the box. For example, we added support for Resources and Prompts as primitives in the MCP protocol in addition to Tools.
Centralized discovery: No more hunting for MCP servers and tools. Agent Registry offers a unified directory where you can easily find and manage agents, MCP servers, and tools in one place.
Easy access with security and governance: Every Google Cloud service is now MCP-enabled by default, allowing agents to easily communicate with Google Cloud. Leverage native Cloud IAM Deny policy for fine-grained access control.
Content safety: With Model Armor in-line integration, you can actively defend against indirect prompt injections and data exfiltration.
Full observability: Maintain full oversight with OTel Tracing and Cloud Audit Logs, helping you troubleshoot your agents, get comprehensive analytics, and perform forensic audits of agentic actions.

Case study: Insta360 redefines video editing

Insta360, one of the world’s leading smart imaging brands, is leveraging the Google Cloud agentic ecosystem to redefine how users capture and share their lives. Building on its "Moments" feature, which delivers AI video highlights, the company developed an AI video editing agent using Google's Agent Development Kit, Agent Engine, A2A and Google-managed MCP servers, allowing users to complete video editing in the cloud through natural language input.

“Transitioning to managed MCP servers will allow us to move away from fragile point-to-point connections and toward a secure, scalable service-oriented architecture. By exposing our proprietary editing tools as managed endpoints, we're gaining the enterprise-grade stability needed to bring autonomous video creation to users around the world.” - Even Lin, Head of Cloud Services, Insta360

Broad coverage across the Google ecosystem

Whether you’re automating internal operations or building customer-facing experiences, Google-managed MCP servers let your models do more than just chat; they provide the secure connections needed to take direct action across your Google Cloud services. Here are a few ways you can leverage them today:

1. Infrastructure, operations, and security

Agents can help move beyond simple monitoring to active orchestration, handling maintenance and monitoring while prioritizing critical security events. Here are a few examples of what this looks like in practice:

Automated lifecycle management (ALM): Provision and decommission resources dynamically based on real-time application demand using GKE, Cloud Run, or GCE MCP servers.
Self-healing systems: Empower agents to monitor events via Cloud Logging and Monitoring, triggering recovery actions like traffic rerouting or deployment rollbacks before users are impacted.
Security orchestration: Build sophisticated workflows that use Google Security Operations to automatically investigate and respond to emerging threats.
Fleet and network ops: Use the Android Management API MCP server to query device health in natural language, or utilize Network Management API MCP server to automate complex diagnostic workflows and surface actionable insights instantly.

2. Databases, analytics, and storage

To be effective, an agent must be grounded in enterprise truth — the live, operational data residing in your production systems. This enables agents to interact with your data ecosystem with:

Real-time operational insights: Connect agents directly to Spanner, AlloyDB, Cloud SQL, Firestore, or Bigtable MCP servers to give agents access to both structured and unstructured operational data, to power immediate, data-driven decision-making.
Analytical insights and data pipelines: Leverage BigQuery and Managed Service for Apache Spark MCP servers to process large datasets, using Pub/Sub or Managed Service for Apache Kafka to trigger proactive system alerts.
Contextual retrieval: Access both structured and unstructured data with Google Cloud Storage and Knowledge Catalog MCP servers to provide agents with real-time context required to accurately execute complex tasks.

3. Services and apps

Beyond raw data, agents require geographic context, technical documentation, and productivity tools to operate effectively. Here are ways agents can leverage these tools to perform more sophisticated and helpful tasks:

Developer assistance: The Developer Knowledge API MCP server grounds AI agents in Google's official developer documentation, allowing tools to reference up-to-date code samples and guides to solve complex technical issues as they happen.
Hyper-local contextual intelligence: Use Maps Grounding Lite MCP server to provide agents with trusted Google Maps data. Deliver high-precision responses for routing, weather, and local points of interest, minimizing hallucinations in travel and real-estate applications.
Conversational experience design: The Gemini Enterprise Agent Platform MCP server empowers agents to act as AI supervisors, programmatically managing the full lifecycle of other models, prompts, and endpoints. Additionally, the Customer Experience Agent Studio MCP server enables AI-assisted workflows for building, modifying, and maintaining customer experience agents.
Productivity and commerce: Streamline team collaboration with Workspace MCP Servers for Gmail, Drive, Calendar, People API, and Chat by allowing agents to summarize Gmail threads, draft Docs, manage Calendar invites and facilitate Google Chat workflows to streamline team productivity. Additionally, with Google Pay and Wallet MCP server, you can integrate payments and digital passes into your agentic workflows, allowing AI to assist with onboarding, troubleshooting integrations, and monitoring checkout performance.

By unifying these capabilities with MCP, it’s easier than ever to build agents that don't just chat, but actually take actions on your behalf.

Technical scenario with demo

To showcase the power of this platform, we built the Pet Passport demo using Google-managed MCP servers and ADK.

This demo features an autonomous agent that handles a complete end-to-end workflow: planning the perfect pet-friendly day out in New York City. The agent implements a macro-to-micro reasoning chain:

First, it uses the BigQuery MCP server to analyze the NYC Dog License dataset and identify the specific neighborhood where a user's dog breed is most popular.
Next, it passes that context to the Google Maps MCP server to generate a verified walking route complete with dog parks and cafes.

When it's time to go live, Gemini CLI leverages the Cloud Run MCP server to deploy the agent application directly from a local machine, publishing the entire experience as a live, shareable URL.

Try it yourself with the codelab or explore the source code on GitHub.

Start building today

We invite you to build production-grade workflows using the Gemini CLI, ADK, or your preferred IDE.

AI & Machine Learning

Gemini Live Agent Challenge: Announcing the winners and highlights

Celebrating our category winners at Google Cloud Next ‘26

Winners

Keep the momentum going

Cloud CISO Perspectives: How Google + Wiz changes multicloud strategy for CISOs

How Google + Wiz changes multicloud strategy for CISOs

In case you missed it

Threat Intelligence news

Now hear this: Podcasts from Google Cloud

Google Named a Leader in the Gartner® Magic Quadrant™ for AI Application Development Platforms: Mid-cycle update

The power of LLMs on your data, more than two orders of magnitude faster and cheaper

Why Proxy Models Work Accurately at Ultra Low Latency and Cost?

How Proxy Models Work?

Revisiting when Proxy Models Work

Why isn’t Vector Search Enough?

Experimental Results

Conclusion

Beyond source code: The files AI coding agents trust — and attackers exploit

Attack surface: What executes

Attack surface: What instructs

Attack surface: What connects

Attack surface: What extends

Applying VirusTotal Code Insight in agentic threat intelligence

Rethinking detection for the agentic era

How Imgix processes 8 billion images daily with G4 VMs powered by NVIDIA Blackwell

The challenge: Instant visuals at scale

Adopting the system that runs Google

Inside the Imgix architecture

Advanced video and image intelligence

The results: 50% faster and up to 6x more throughput

Orchestrating at scale

The future: From experimentation to execution

Get started

Google Cloud and NVIDIA expand AI innovation across industries at GTC 2026

Cloud Storage Rapid: Turbocharged object storage for AI and analytics

Rapid Bucket

Rapid Cache

Choose your rocket ship

Cluster-level reliability for trillion-parameter models on TPUs

Reliability for AI supercomputers

Deep dive: The mathematics of availability at scale

Scale of modern AI hardware

Enhancing ML productivity

Enabling the next generation of AI breakthroughs

Gemini 3.1 Flash-Lite is now generally available on Gemini Enterprise Agent Platform

Software development and engineering

Customer experience and high-volume service

Creative pipelines and gaming

Financial services and data operations

Get started

How BASF manages thousands of supply chain decisions with AlphaEvolve’s agentic algorithms

Planning across a two-year lead time

Building a foundation for decision support

Measuring what good looks like in initial tests

The results

What the evolved algorithm actually does

What's next

Fitting the future: How Breuninger boosted sales with its "be your own model" AI

From trusted tester to live product

Three levels of virtual try-on

Building for scale with Flutter

Real results during the holiday season

What’s next

Pioneering AI-assisted code migration: How Google achieved 6x faster migration from TensorFlow to JAX

Accelerating the transition from TF to JAX

Moving beyond single-agent coding

Scalable validation and dynamic Playbooks

Redefining migration velocity

Looking ahead into the AI-assisted era

The Blueprint: Translating stream-of-conscious speech into responsive, actionable task lists

The challenge:

The solution:

The architecture:

The outcome:

The details:

Five must-have guides to move agents into production with Gemini Enterprise Agent Platform

1. Agent design patterns for long-running AI agents

2. The agent governance stack

3. Must-have multi-agent orchestration patterns in ADK