AI & Machine Learning

The Blueprint: How Voicify makes AI-enabled ordering a delight for customers

Thu, 23 Jul 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 2018, Voicify reimagines the traditional phone call with the goal of transforming every call into a seamless and engaging experience.

The challenge:

When we started Voicify in 2018, our vision was to help organizations build confident, pragmatic, and technically grounded voice-driven assistants for any channel, including phones and chat. But the pandemic changed everything.

We shifted our focus to telephone use cases primarily in the restaurant and healthcare sectors where, at the time, call volume and staffing posed significant challenges. Restaurants could potentially miss up to 20% of their calls and lose orders as a result, and healthcare providers struggled to keep up with call volume with the required 100% accuracy when integrating appointment information into a practice management system. We realized that specialized, purpose-driven AI assistants were the key to businesses maintaining excellent service at scale.

To succeed, we had to overcome four primary challenges:

Transactional precision: Our voice assistant needed to reason with complex customer requests against point of sale and practice management systems with 100% accuracy.
Traffic spike management: Our LLM usage needs to be provisioned accurately to keep costs down and maintain customer services in spite of the common (and extreme) spikes in traffic seen in restaurants and healthcare organizations.
Latency: Any delay in the assistant’s response can cause customers to hang up. We needed superfast time to first token, with minimal delay from when a user sends a voice or text request to when the AI model generates its first piece of output.
Security and compliance: Since our founding in 2018, we’ve ensured that we’re HIPAA, SOC2, ISO27001, and PCI-compliant, and that our security is enterprise-grade. We needed architecture and infrastructure that employs all possible safeguards to safeguard data integrity and security.

The solution:

Our conversational orchestration platform builds and validates restaurant orders against a point-of-sale system before submission to ensure accuracy. Under the hood, Gemini Flash, served via Gemini Enterprise Agent Platform, vastly improves latency, minimizing user wait times and preventing hang-ups. With it, we also see approximately 25% to 30% savings compared to our previous use of other LLMs, and with greater reliability too.

To grow the business — and call volume — and to handle traffic spikes, we switched from Google AI Studio to Vertex AI and its current incarnation in Gemini Enterprise. We wanted the enterprise guarantees the latter provided, which we needed for scaling as well as for security and compliance for our healthcare clients. Specific Gemini Enterprise Agent Platform features help us manage high call volumes without experiencing service interruption or dropped responses.

These enterprise-grade services may have carried an increased cost over AI Studio, but they were well worth it to ensure reliable uptime, and the premium pay-as-you-go feature made scaling much easier for us. For example, we used a combination of provisioned throughput and premium pay-as-you-go with Vertex AI to accommodate all-time high usage the day before Thanksgiving, and we saw no rate limiting issues.

The architecture:

The outcome:

Gemini has reduced the burden on our in-house programmatic tools for pulling context and building menus. We’ve seen great improvements in performance and reliability, with lower latency and greater reliability with Gemini. And, the increased stability of our Gemini-powered assistants has made client onboarding much more efficient. Now it only takes one to two days to get a restaurant ready to test after gaining access to the POS system, down from what previously took one to two weeks.

With Google solutions for scale and enterprise-grade service, we’ve optimized our critical time-to-first-token metric, minimizing customer wait times. Using Vertex AI’s provisioned throughput and pay-as-you-go features, we’ve ensured 100% uptime, prevented dropped responses and rate-limiting issues, even during periods of all-time high usage. We’re now able to easily manage the spiky nature of restaurant traffic.

In terms of technology, we anticipate moving beyond conversational order capture to more proactive assistance, using context from conversations or POS activities. Your typical Friday night order from your favorite Japanese restaurant? Someday soon it might be Voicify’s voice assistant proactively placing it for you.

The details:

Our industry focus presents a few unique challenges that we had to spend time solving within the backend.

The core component of our Voicify solutions is our voice orchestration platform, which manages the entire phone AI stack and is designed for enterprise-grade scalability and security. This is also the node where industry solutions are called depending on user needs.Our voice orchestration platform sits close to the customer and coordinates backend services like Gemini and the different components of the voice assistant. We use it to manage functions like automated speech recognition, text-to-speech, and text generation, which is not purely generative but includes programmatic elements.

One of the unique architectural decisions we made was how to manage large, complex restaurant menus. We decided to avoid putting the entire menu into a single prompt, and we include only certain information in the initial prompt and then gather more details as the conversation progresses. This improves response times and helps manage the complexity of larger orders by focusing on only the relevant parts of each menu in a given interaction.

We also designed the architecture from the outset of our company to meet the high standards of enterprise clients for security and compliance, particularly in healthcare. We are making sure that our scalability is enterprise-grade. Architecturally we’re also employing all safeguards to ensure data integrity and safety too. Lastly, our platform is designed to support a multicloud environment as part of our strategy for achieving the highest possible level of availability.

Why AI apps fail in production (And how Google solved it)

Tue, 21 Jul 2026 23:00:00 +0000

We are living in the golden age of the weekend AI side project. Thanks to agentic engineering and LLMs, the time to go from a blank IDE to a functional local application has dropped from quarters to hours. You can build your wildest ideas over a cup of coffee.

But inside an enterprise ecosystem with rigid infrastructure and millions of users, vibe coding hits an invisible wall. Your local prototype falls apart against corporate networks, cascading errors, or getting blocked by leadership terrified of operational volatility.

The data is sobering: only 5% of AI prototypes make it to production; the other 95% fall into the validation abyss.

For developers, watching people on social media ship lightning-fast AI deployments while you’re stuck in endless validation loops is maddening. To figure out how to bridge this chasm, I went into the engineering trenches at YouTube to see how they manage this exact speed-versus-risk paradox. What I discovered completely rewrites the playbook on AI software development lifecycle (SDLC) design.

The risk-vs-speed paradox

When you are solo-building, failure is cheap. Writing agentic code is like piloting a nimble jet fighter—if an AI agent misbehaves, you rewrite the prompt and instantly restart the server.

But as AI engineering leader Addy Osmani points out in our premiere of Emergent, unconstrained agentic orchestration inside an enterprise introduces an unpredictable blast radius. Addy recalls running ten parallel agents on a personal project, context-hopping and pushing code based purely on quick previews. The technical debt accumulated fast, breaking two apps catastrophically because the modifications weren't properly isolated.

Amplify that risk to the scale of YouTube. Its infrastructure handles billions of users on a robust, 20-year-old codebase. It is essentially a public utility; you cannot risk overloading it with experimental technical debt. Protecting a platform of this scale requires extensive, slow guardrails:

By the time you build a primitive demo through this pipeline, the underlying AI models have evolved, leaving your idea out of date. How do you move at lightspeed while minimizing systemic risk?

YouTube’s AI prototyping stack

Deepmind and former YouTube software engineer, Benji Bear, solved this puzzle not by accelerating reviews, but by changing infrastructure philosophy. He and his team built a prototyping stack — a unified design-to-code lifecycle platform that completely decouples rapid experimentation from mainline production servers. It systematically solves the two primary friction points of developer velocity.

Decoupling the data layer

Isolating a standalone app completely causes a "blank canvas" problem where you can't test prototypes against realistic conditions. To solve this, developers bootstrap their ideas using pre-built Google AI Studio templates. These templates hook into a proxy server set up on Google Cloud for prototype-approved read-only data. This instantly grants the prototype pre-authenticated, read-only API access to live metadata bundles (playlists, videos, channels) via strict tokens.

Developers get the technical accuracy of live production parameters without any ability to write back to, pollute, or crash core databases.

Live UI injection

When a concept requires true real-world validation, the stack offers client-side YouTube Extension wrappers. This wrapper acts as glue code, allowing developers to inject their experimental features directly into the actual, live production web surface of YouTube. Code-split chunk safeguards isolate this from production binaries, allowing prototype updates to deploy to a safe staging environment in minutes.

The result? YouTube went from taking multiple quarters to vet an idea to launching several successful prototypes — including YouTube Recap and Ask YouTube — straight to user research studies (UXR) in weeks.

Embrace throw-away code

Implementing this stack requires a profound psychological shift. Engineers are trained to treat code as permanent infrastructure, polishing and refactoring it until it’s pristine. But Benji’s core enterprise AI philosophy here is simple: Embrace throw-away code.

Google AI Studio prototypes are meant to be messy with some technical debt; their objective is to validate product-market fit using quantitative data. Trying to refactor a chaotic, AI-generated app into an enterprise codebase is an architectural trap that can create friction.

But because Google AI Studio builds your prototype directly onto a mirrored version of production infrastructure, you establish a highly accurate baseline from day one. You still discard the messy, AI-generated script, but when an idea proves successful, rewriting it for production becomes significantly faster, cheaper, and safely positioned later in the development lifecycle—giving you a verified blueprint to code against rather than a blank canvas.

Move fast without breaking things

The core realization here is that a 95% failure rate isn’t a bug — it is the strategy. We should design environments that encourage our teams to fail more frequently and safely.

AI has plummeted the cost of code generation. Consequently, our roles are shifting from syntax gatekeepers to system architects. Our job is to design the bridges, read-only sandboxes, and isolated pipelines that empower teams to test wild ideas without triggering catastrophic meltdowns.

The biggest risk isn't breaking a server with messy AI code; it's missing the technological moment because validation loops are too slow. By building structural constraints that make failure safe, you give your team the freedom to run at hyper-speed.

To see the full technical breakdown, interview clips with YouTube's core infrastructure engineers, and a look inside the Google AI Studio Proto-Stack, watch our premiere episode of Emergent on YouTube.

Now in preview: Find and fix software vulnerabilities with CodeMender

Tue, 21 Jul 2026 15:00:00 +0000

As adversarial AI threats accelerate attacks on code, security teams must counter them with machine-speed defenses that can automate code remediation and fight AI with AI.

CodeMender is our managed code security agent, and starting today, we're bringing its code scanning and remediation capabilities directly to you in preview.

CodeMender offers access to our generally available models via Gemini Enterprise Agent Platform, or it can be deployed as a core component of AI Threat Defense.

CodeMender also aligns with our multi-model approach, so you can choose the right model to optimize for cost, speed, and deep scanning performance. It will support third-party frontier model options later this year.

Watch this overview of CodeMender in Gemini Enterprise Agent Platform.

CodeMender can help you advance from passive scanning to automated code remediation, and reduce zero-day risk. It examines and remediates existing code security issues without sacrificing development velocity by:

Deploying the best-fit model. You can choose from multiple models to optimize for costs, speed, deep scanning, and coding performance.
Automating machine-scale remediation. You can now eliminate remediation bottlenecks caused by manual verification and patching, while keeping developers in the loop.
Prioritizing fixes by exploitability. You can run proof-of-concept exploits and execute simulations to verify that vulnerabilities in the code are exploitable, and prioritize resources on fixing the most critical issues first.

Find and fix vulnerabilities with AI

Born from Google DeepMind's pioneering AI research, CodeMender transforms vulnerability management from a manual bottleneck into an autonomous, high-speed system. Your developers and security practitioners can automatically scan software for flaws, verify them with executable exploits, and remediate them with tested code fixes.

“At Salesforce, trust is our number one value, and protecting customer data means continually raising the bar for how we find, validate, and mitigate risks. CodeMender brings AI into a critical part of the security lifecycle by accelerating the path from validated vulnerability to tested fix. As AI reshapes the threat landscape, capabilities like this help strengthen resilience and give our customers the confidence to keep innovating,” said Iain Mulholland, CISO, Salesforce.

"CodeMender consistently identified critical vulnerabilities that our other AI-enabled tools completely missed. It doesn't just find theoretical flaws — it proves the immediate risk and delivers targeted, validated fixes that secure our environment without disrupting core business logic," said Scott Ponte, head, Security Operations, Robinhood.

"CodeMender is fast, comprehensive, and genuinely ambitious about closing the loop from detection to fix, enabling teams to secure their software supply chain without losing velocity," said Ashwin Kannan, principal AI engineer, Office of the CTO, Palo Alto Networks.

How the CodeMender agent works

We’ve fine-tuned CodeMender’s harness to be continuously updated with the latest Google DeepMind research, including the up-to-date agent skills, security tools, and system prompts.

Operating in the secure-by-design Agent Platform, CodeMender is protected by enterprise-grade, built-in governance and security guardrails, including secure traffic routing through your VPC, data isolation and encryption, and zero retention of source code data.

As an agent, it can integrate with existing continuous integration and continuous delivery (CI/CD) workflows, or run directly in local developer environments using a lightweight command-line interface (CLI) client.

You can also configure CodeMender to scan and analyze code in a sandbox that you manage. The agent connects to your code repositories and works with developer tools, such as VS Code and Antigravity, to safely analyze first-party, open-source, and third-party software.

Scan: Find hidden vulnerabilities with flexible model scanning

CodeMender scans for top vulnerability classes and understands the unique context, goals, and functionality of your software repositories and applications.

Scan: Discovered new vulnerabilities and categorized by severity and type.

CodeMender’s harness with security context helps you discover sophisticated vulnerabilities that static and model-only scanning miss. These scans look for hard-to-find vulnerabilities like memory corruption, injection, web security issues, cryptographic flaws, and insecure data handling. CodeMender supports common software languages including C/C++, Go, Java, Python, Ruby, Rust, and TypeScript.

Verify: Simulate and verify exploits to reduce noise

CodeMender can help cut alert fatigue and false positives by proving a vulnerability presents a legitimate risk before fixing it. The agent goes beyond static code-pattern analysis by simulating an attack with exploit code it builds and runs in an isolated, customer-managed sandbox.

Verify: Creates verification plan and builds and tests exploits in your sandbox environment.

The agent uses this proof-of-concept exploit to verify that the security flaw poses a legitimate risk. This critical verification phase allows your security practitioners and developers to prioritize validated risks by eliminating false positives.

Remediate: Automatically generate and test code fixes

Identifying risky security flaws is only half the battle. Once a vulnerability is verified, CodeMender automatically generates a secure patch to resolve the issue. The fix is delivered as a code difference directly in developer tools, so it can be integrated into existing development workflows.

Remediate: Generates and tests code fix with code diff for developer review and approval.

CodeMender further strengthens the fix by using LLM-as-a-judge to ensure it doesn’t disrupt existing application functionality. You can even provide context on your codebase's distinct coding conventions and styles so that CodeMender generates code that matches it. Developers remain in full control, manually reviewing and approving CodeMender's patches before any code is committed to the repository.

CodeMender in AI Threat Defense

When leveraged as part of AI Threat Defense, Wiz orchestrates agentic application security, analyzing applications to prioritize investigations. It calls CodeMender to scan code (coming soon), enrich findings within the Wiz Security Graph with deployment context, and trigger Wiz Red Agent for AI pentesting to prove exploitability, ensuring that teams focus on the highest-risk vulnerabilities.

Through Wiz, AI Threat Defense calls CodeMender to scan code, enrich findings, and trigger AI pentesting.

Wiz serves as a command center for governing and scaling remediation in AI Threat Defense. The Wiz Green Agent orchestrates this lifecycle by directing CodeMender to generate and test high-fidelity patches enriched with application context from the Security Graph. This workflow empowers teams to resolve complex vulnerabilities with unprecedented speed and precision.

How to get started with CodeMender

Consistent with our multi-model approach, CodeMender can help you optimize for cost, speed, and deep scanning performance.

You can use CodeMender with our generally available Gemini models via Agent Platform, or deploy it as a core component of AI Threat Defense.

Separately, CodeMender with Gemini 3.5 Flash Cyber will be exclusively available to a small set of governments and trusted partners. We plan to expand this access over time.

CodeMender is a critical step towards a continuous, self-healing agentic software development lifecycle, a future where code is autonomously secured, validated, and patched before it ever hits production.

You can learn more about CodeMender and review the documentation here.

13 hands-on demos to build on Gemini Enterprise Agent Platform

Fri, 17 Jul 2026 16:00:00 +0000

Earlier this year, we introduced Gemini Enterprise Agent Platform, where you can build, scale, govern, and optimize agents. Today, we’re sharing 13 demos that walk you through what Agent Platform can do. Each one teaches a concept, a pattern, or an architecture you can put to work immediately.

The best part? You don't have to follow them step-by-step. Install Agents CLI into your favorite coding agent (Antigravity, Claude Code, Codex, whatever you use) and it instantly gets seven skills that make it an expert in ADK and Agent Platform. Describe what you want to build in plain English, and your coding agent scaffolds, evaluates, deploys, and monitors the agent for you. You’ll never have to leave your editor.

Let’s dive in!

Build AI agents

These demos are all built on the code-first ADK. They start at the foundation and work up.

1. Start here: build your first agent with ADK. The ADK Foundation codelab is your perfect on-ramp. You set up your environment, define a basic conversational agent powered by Gemini, configure its settings, and test it through both a command-line interface and a web UI. If you've never touched ADK before, do this one first.

2. Build an event-driven approval agent with human-in-the-loop. The ambient expense agent codelab is the most complete "Agent Platform in action" demo in the set. You build a corporate expense agent using ADK 2.0's graph-based workflow API. Expenses under a threshold get auto-approved in plain Python. Anything above goes through a pre-LLM security screen (PII redaction, prompt-injection defense), passes a Gemini compliance analysis, and pauses for a human-in-the-loop review before anything is finalized. You mount it behind FastAPI, trigger it from Pub/Sub events, and grade it with an LLM-as-judge eval. Keep this agent in mind – it comes back in the Scale and Govern sections.

3. Connect agents to your data with the Model Context Protocol. The MCP codelab shows you how to build reusable MCP tools that let Gemini query BigQuery, search files, and call APIs. MCP is an open protocol, so the tools you build work across different vendors and frameworks.

4. Build a dynamic frontend with Agent-to-UI (A2UI). The best user experiences are highly visual. The A2UI codelab shows you how to build an agent that renders real interface components (layouts, charts, interactive menus) that update dynamically in real time as the conversation flows. The agent literally assembles the UI the user needs, on the fly.

Scale AI agents

A prototype on your laptop is one thing. Handling production traffic, memory, and orchestration is what comes next.

5. Deploy a stateful data science agent to Agent Runtime (formerly known as Agent Engine). The Stateful Data Science Agent codelab walks you through building a BigQuery agent that remembers user preferences across sessions via Memory Bank, then deploying it directly to Agent Runtime. All of the underlying infrastructure, scaling, and session management are handled for you automatically.

6. Build long-running agents that pause, resume, and never lose context. Building an agent that responds to a single prompt is easy, but real enterprise workflows often take days or weeks to complete. This tutorial walks through building agents that run reliably for weeks. You'll learn three architectural patterns: durable state machines, event-driven idle time handling, and checkpoint-and-resume with persistent sessions. The example is an onboarding coordinator agent that survives container restarts and picks up exactly where it left off.

7. Deploy an ambient expense agent to Agent Runtime with the Agents CLI. Remember the expense agent from the Build section? The Deploy to Agent Runtime codelab picks up that agent and takes it to production. You scaffold your deployment config with the Agents CLI, preview it with a dry run, then deploy it live. Cloud Trace, Cloud Logging, and BigQuery Agent Analytics wire in automatically, and the agent auto-registers in Agent Registry, so it’s discoverable across your org the moment it goes live.

8. Give your production agent a real front end. The frontend codelab is where everything comes together. You build a manager dashboard on Cloud Run, connect it to Agent Runtime through an OIDC-authenticated Pub/Sub pipeline, and give managers the ability to resume paused human-in-the-loop sessions from the browser. It ties the expense agent and the deployment together into a complete end-to-end enterprise architecture.

Govern AI agents

Scaling agents across an organization requires a system of built-in guardrails to manage access, track endpoints, and filter traffic.

9. Secure your agent's lifecycle from the first commit. The Secure Agentic Coding codelab shows you how to build a shopping assistant test-first with test-driven development (TDD), wire in a custom STRIDE threat model, set up a Semgrep pre-commit hook, and configure a PreToolUse gate that blocks risky actions before execution. You deliberately plant a hardcoded API key, and the agent catches and fixes it the moment the hook fires.

10. Control agent access with Agent Gateway. The Agent Gateway codelab covers runtime governance. You deploy a multi-tool ADK agent on Agent Runtime that calls MCP servers on Cloud Run through Agent Gateway. Each agent gets a unique identity with end-to-end mTLS. Every outbound call goes through IAP authentication and IAM authorization. On top of that, Model Armor inspects all content for prompt injection and data leakage. It’s a complete, production-grade governance stack in one demo.

Optimize AI agents

Shipping an agent is the start. The hard part is knowing whether your next prompt tweak actually makes it better or quietly breaks ten other things. Agent Platform gives you the tools to close that loop.

11. Drive the agent quality flywheel from your coding agent. You tweaked a prompt. It looks better on three examples, but did you just break ten others? This tutorial introduces a five-stage evaluation flywheel you run directly from your coding agent: prepare data (from OTel traces, hand-crafted cases, or synthesized scenarios), run inference, grade with Google's adaptive AutoRaters, analyze failure clusters, and execute targeted optimizations. The AutoRaters are built on the same principles Google uses to evaluate its own models and first-party agents, developed in partnership with DeepMind. Describe what you want measured in plain language. Your coding agent picks up the rest.

12. Build a cross-language multi-agent pipeline with A2A. In a large enterprise, different teams will inevitably build agents in different languages. This tutorial walks through a contract compliance pipeline where a Python-based agent extracts terms using Gemini and a Go-based agent validates them against corporate policy. The two services connect via the Agent-to-Agent (A2A) protocol and are orchestrated by ADK. You'll learn how RemoteA2aAgent turns any A2A-compliant service into a local sub-agent with a few lines of code.

13. Scale agents across frameworks with CrewAI, LangGraph, A2A, and ADK. Most production teams don't standardize on one agent framework. The codelab shows you how to orchestrate across all of them: an ADK control room delegates planning to a LangGraph state machine, which dispatches tasks to a CrewAI execution crew, all connected via the A2A protocol. If one step fails, the control room re-plans automatically.

Get started

If you want to see the full agent development lifecycle in under 10 minutes, watch this walkthrough. Otherwise, install Agents CLI, open up your coding agent, and start building today.

Google is a Leader and positioned furthest in Vision and highest in Execution in the 2026 Gartner® Magic Quadrant™ for Conversational AI Platforms

Thu, 16 Jul 2026 19:00:00 +0000

For the second consecutive year, Google has been named a Leader in the Gartner® Magic Quadrant™ for Conversational AI Platforms. Google received the furthest and highest in positioning on the "Vision" and "Execution" axes and is now ranked #1 in three out of four Critical Capabilities Use Cases. We believe this recognition reflects our continued investment in frontier AI research, enterprise infrastructure, and helping customers move AI from experimentation into production at scale.

More importantly, we believe it reflects the success of the organizations building with Gemini Enterprise for Customer Experience every day.

Figure 1: Magic Quadrant for Conversational AI Platforms (Image of the Gartner Magic Quadrant for Conversational AI Platforms, showing Google positioned in the "Leaders" quadrant.)

Download the complimentary 2026 Gartner Magic Quadrant for Conversational AI Platforms.

Building the next generation of customer experiences with Gemini Enterprise for Customer Experience

Enterprise customer experiences are entering a new era. Organizations are moving beyond traditional chatbots toward AI agents that can understand customer intent, reason across enterprise knowledge, and take action across business systems.

As these experiences move into production, enterprises need more than powerful models. They need an AI platform that combines frontier research with enterprise security, governance, operational reliability, and the ability to scale globally.

Today, Gemini Enterprise for Customer Experience brings these capabilities together to give your customers a frictionless experience. Organizations can deploy agents that eliminate disjointed interactions across voice and digital channels, allowing customers to discover, purchase, and get help across every touchpoint without starting over. This connected journey drives revenue growth, deeper loyalty, and lower operational costs.

Built for production AI

At the center of Gemini Enterprise for Customer Experience is CX Agent Studio, Google’s platform for building intelligent customer experience agents. By coupling our newest models, unified product capabilities, and updated deployment best practices, we abstract technical complexities so enterprise teams can build at an unprecedented speed and derive true business value.

Organizations can use CX Agent Studio to:

Build multimodal AI agents and deploy them across voice and chat channels,
Assist human support and service representatives in real time,
Analyze customer conversations to improve business outcomes,
And, accelerate deployment with pre-built agents for industries including retail, food ordering, and automotive.

Modern customer experiences demand more than answering questions. They require AI that can understand complex requests, retrieve trusted information, reason through multiple steps, and take action across enterprise systems. For example, The Home Depot is already using these capabilities for customer support - helping customers reach solutions up to 4x faster than traditional phone menus when calling into a store. AI voice agents built with CX Agent Studio understand why a customer is calling in fewer than 10 seconds to help customers complete purchases, initiate service requests, or seamlessly transition to a human associate when needed.

“AI does a tremendous job at recognizing customer intent and taking direct action to help complete a purchase or even start a service request. And of course, if they need to speak with an associate, we’ll quickly connect them.” - Jordan Broggi, EVP of Customer Experience and President of Online, The Home Depot

CX Agent Studio combines native multimodal capabilities, agent orchestration, enterprise retrieval, and integrated developer tooling to help organizations move quickly from experimentation to production.

Whether deploying pre-built industry agents or building custom experiences, organizations maintain enterprise-grade security, governance, and operational controls while retaining complete ownership of their customer experience.

Powered by Google’s AI optimized stack

Gemini Enterprise for Customer Experience is built on Gemini models developed by Google DeepMind. But having access to Google DeepMind's world-leading research and frontier models is the starting line. A brilliant model is only as powerful as the foundation it runs on. To put human-grade customer experience agents into production - where milliseconds of latency matter for voice interactions and hallucinations pose real business risks - you need a platform engineered for performance.

This is why Gemini Enterprise for Customer Experience and CX Agent Studio run natively on Google Cloud’s complete, first-party AI stack. Spanning from our custom-built AI infrastructure (AI Hypercomputer) and the Agentic Data Cloud that grounds your models in real-time truth, up to the autonomous protection of Agentic Defense, every layer is co-designed to function as a single, unified system on a foundation of uncompromising security.

For enterprise CX leaders, this is your structural edge. Because your agents are built on this unified stack, they automatically benefit from our continuous advancements - absorbing every new DeepMind capability and hardware efficiency we achieve. This deep integration delivers the speed, safety, and cost-efficiency you need, freeing your teams to focus on building the next generation of customer experiences.

Looking ahead

The next generation of customer experiences won’t simply answer questions. They’ll understand context, reason across enterprise knowledge, collaborate with people, and take meaningful action on behalf of customers.

Our vision is to help organizations build AI agents that are proactive, personalized, and continuously improving across every customer touchpoint.

To download the full 2026 Gartner® Magic Quadrant™ for Conversational AI Platforms report, click here. For more information on CX Agent Studio and Gemini Enterprise for Customer Experience, visit our website.

_{Gartner, Magic Quadrant for Conversational AI Platforms, Gabriele Rigon, Justin Tung, Arup Roy, Adrian Lee, Uma Challa, July 7, 2026}

_{Gartner, Critical Capabilities for Conversational AI Platforms, Justin Tung, Uma Challa, Adrian Lee, Gabriele Rigon, Arup Roy, July 7, 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.}

Cloud CISO Perspectives: How AI leverages deep context as the defender’s advantage

Thu, 16 Jul 2026 16:00:00 +0000

Welcome to the first Cloud CISO Perspectives for July 2026. Today, Francis deSouza, COO, Google Cloud and President, Security Products, explains the crucial role that deep context plays in creating an AI advantage for defenders.

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 0x7f814210d820>), ('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 AI leverages deep context as the defender’s advantage

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

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

Attackers are making headlines with AI, but defenders have a distinct and powerful advantage.

AI is rapidly transforming the cyberthreat landscape, driving unprecedented shifts in the scale, speed, and sophistication of attacks. Just recently, Google Threat Intelligence Group documented a critical milestone: the first known case of a zero-day exploit built entirely with AI. While we successfully disrupted their plans and got the vulnerability patched before launch, it highlights exactly what we are up against.

With AI agents, attacks are accelerating at machine speed. The handoff time between the first and second stage of an attack used to be eight hours; today, it takes just 22 seconds.

There’s an old saying in cybersecurity that adversaries only have to be right once, but defenders have to be right every time. That is the attacker’s advantage.

But AI is rewriting those rules, delivering a decisive defender's advantage built on deep context.

The AI Era: Attacker’s Profile vs. Defender’s Advantage

Aspect	Attacker's Profile	Defender's Advantage
Visibility	Limited to outside-in probing; little enterprise context upon entry.	Complete inside-out context; knows exact asset locations, application behavior, and team ownership.
Operational Speed	Executes multi-agent handoffs in 22 seconds.	Machine-speed defense; proactive mitigation in seconds (such as Morgan Stanley's 90-second resolution.)
Core Tactics	Multi-model phishing, deepfakes, AI-built zero-days, and model poisoning.	Closed-loop defense; continuous exposure mapping and accelerated code patching.

The unified blueprint: Google AI Threat Defense

Previously, enterprise context data was fragmented across disconnected security tools. Now, AI empowers defenders to synthesize this rich data into a unified, always-on, autonomous defense.

We built Google AI Threat Defense to combine Google’s security capabilities into a single platform: the advanced reasoning of Gemini, the contextual cloud power of Wiz, the code-level remediation capabilities of CodeMender, and the frontline intelligence of Mandiant.

Our platform transforms vulnerability management across a continuous four-step framework:

Stage	Technology & Actions	Strategic Value to the Enterprise
1. Prepare	Map exposed applications, APIs, identities, and runtime environments using Wiz. Simulate attack paths with the Wiz Red Agent.	Hardens the foundation to reduce internet reachability before vulnerabilities hit production.
2. Scan & Prioritize	Run multi-model scanning — using lighter models for broad coverage and Gemini frontier models for deep-dive analysis of high-risk assets.	Replaces massive alert lists with deep, context-driven risk validation, including an optimal cost per token.
3. Remediate	Deploy CodeMender inside developer IDEs/CLIs to auto-generate verified code fixes.	Replaces slow, manual patching with autonomous code-level remediation and memory-safe migrations.
4. Monitor	Deploy AI agents tied to Wiz to hunt for vulnerabilities and anomalies across network, identity, and application telemetry. Pair with Google Security Operations to rapidly hunt for unknown threats.	Establishes machine-speed runtime detection for zero-day response and threats against unpatchable environments.

To stop vulnerabilities before they hit production, Morgan Stanley partnered with Google Cloud and Wiz, aligning their strategy with the core principles of the AI Threat Defense framework: prepare, scan, remediate, and monitor. By replacing fragmented tools with this unified blueprint, Morgan Stanley collapsed its mean time to detect threats by 99.9%, shifting from a reactive 45-minute window to proactive mitigation in 90 seconds or less.

Google Cloud x Morgan Stanley: Redefining Threat Defense in the AI Era

Maintaining strategic human oversight

While human-speed execution cannot keep pace with automated threats, human management remains essential. We align autonomous AI agents directly with the human teams they support. In Wiz, for example, the Red agent automates penetration testing, the Blue agent drives threat investigations, and the Green agent accelerates cloud remediation.

Every AI conversation is a security conversation. That means securing AI infrastructure requires building from the ground up, and not bolting on.

This ensures autonomy under human supervision, empowering engineering and security teams to eliminate backlogs and secure the software development lifecycle without sacrificing speed.

What’s next: AI-native, agent-driven infrastructure

The foundation of your defender's advantage starts with protecting your environments — not just from outside threats, but from internal risks like shadow AI and unauthorized agents. When employees download models and deploy agents outside of IT oversight, they create silent logic breaches and data-poisoning risks.

The key to countering this is enforcing Zero Trust for AI, and directing teams toward approved architectures with proper governance. Every AI conversation is a security conversation. That means securing AI infrastructure requires building from the ground up, and not bolting on.

At Google, security is not just an added layer; it is our foundation. Our secure-by-default architecture automatically blocks nearly 15 billion unwanted emails and protects billions of users every day.

As the threat landscape matures, outperforming automated adversaries requires a platform built from the ground up to be AI-native and agent-driven.

Fight AI with AI. Learn more about how to secure your software lifecycle with Google AI Threat Defense.

aside_block: <ListValue: [StructValue([('title', 'Learn something new'), ('body', <wagtail.rich_text.RichText object at 0x7f814210dcd0>), ('btn_text', 'Watch now'), ('href', 'https://www.youtube.com/watch?v=CmGWIwgHR60'), ('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:

FinOps for SecOps: How to optimize the agentic SOC for value: To be more resilient in AI adoption, CISOs should develop a disciplined "FinOps for SecOps" blueprint that maximizes threat disruption while keeping control over compute costs. Here's how. Read more.
New IDC study: How Mandiant transforms security into a competitive advantage: A new IDC Business Value White Paper found that you save an average of $4.3 million, driving a 268% three-year ROI, with Mandiant Consulting. Read more.
Drive proactive security, prioritize risks with Google Threat Intelligence and Wiz ASM: To help you match your real-world exposures with real-time adversary activity, we’ve begun integrating Google Threat Intelligence with Wiz Attack Surface Management. Read more.
Shift into high gear with agents: Securing the software-defined vehicle: To better support and secure SDVs, Google Cloud and Valtech have partnered to develop Nexus SDV, a highly-scalable, AI-enabled connected vehicle platform. Read more.
Meet the 33 cybersecurity startups joining the Gemini Startup Forum: Our flagship Google for Startups program, Gemini Startup Forum: Cybersecurity, has selected its first 33 trailblazing startups. Read more.
Introducing k8s-aibom on GKE for automated AI bills of materials: We’re open-sourcing k8s-aibom, a Kubernetes controller that continuously monitors environments to detect AI runtimes and generate standard ML-BOMs. Read more.
BGP route policies: Top 3 use cases by customer demand: We detail the three most impactful use cases for Cloud Router BGP route policies that have emerged since 2025. Read more.
Contributing to U.K. financial sector resilience as a critical third party: The U.K. Treasury has designated Google Cloud EMEA as a critical third party (CTP) to the U.K. financial sector under the CTP regime. Here’s how that helps you. Read more.
Google Cloud confirmed to offer a safer choice for EU public sector organizations with Dutch DPIA approval: We understand that for the EU public sector, data protection is a prerequisite. We’re excited to reinforce this commitment with a major milestone. Read more.
Why IaC coverage belongs on your security dashboard: Rethinking infrastructure-as-code coverage as a funnel that shows how much of your infrastructure is governed, traceable, and ready for remediation at speed. Read more.
Inside the ProdSec playbook: Operationalizing Wiz for end-to-end cloud security: Rethinking infrastructure-as-code coverage as a funnel that shows how much of your infrastructure is governed, traceable, and ready for remediation at speed. Read more.
Build AI security agents with Wiz MCP: Power AI-driven security with trusted security context, Wiz AI Agents, and Wiz AI Skills. 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 0x7f814210d880>), ('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

A look at the drivers, dynamics, and applications of the pro-Russia influence ecosystem: Four years into Russia’s full-scale invasion of Ukraine, the pro-Russia influence ecosystem has evolved from a tool of war back into a global strategic asset. The interconnected nature of the ecosystem's disparate components makes it resilient to limited scope disruptions, a factor that defenders need to consider to mitigate pro-Russia influence threats. Read more.
Google’s continued disruption of malicious residential proxy networks: In coordination with the FBI, Lumen, and others, Google took action against the NetNut residential proxy network, also known as Popa. This action builds on our disruption of the IPIDEA proxy network that took place in January 2026, and is a continuation of Google’s objective to dismantle malicious residential proxy networks. Read more.
GhostApproval: A trust boundary gap in AI coding assistants: Learn how Wiz uncovered a category-level blind spot in modern AI coding assistants, and why the human-in-the-loop safety model fails against this classic threat. Read more.
The latest addition to Turla’s intelligence gathering apparatus: Google Threat Intelligence Group (GTIG) has conducted an in-depth analysis of a .NET backdoor, tracked as STOCKSTAY, that has been continually developed and deployed by the Russia-linked threat actor Turla, one of the oldest known cyber espionage groups, since at least December 2022. As part of our continued tracking of this group, we’re providing an overview of our STOCKSTAY analysis, a timeline of key developmental and operational observations, and detailed similarities to KAZUAR to contextualize this new capability in Turla’s arsenal. Read more.
Recovering active ADFS signing keys via Machine DPAPI: During a recent red team engagement, Mandiant discovered that when ADFS certificates are manually rotated, configuration drift can silently leave active signing keys exposed in Machine DPAPI. Here’s how to defend against it. Read more.

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

Now hear this: Podcasts from Google Cloud

Cloud Security Podcast: Building an AI-pilled, solo vibe-coded, Clickhouse-based SIEM: Dan Lussier, founder, Nano, unpacks how he vibe-coded an entire SIEM from scratch during his end-of-year holiday break. Listen here.
Cloud Security Podcast: Scaling lessons, from leading the NSA to defending the world: Morgan Adamski discusses how public-private partnerships and the shift to cloud infrastructure have transformed cybersecurity defense through improved intelligence sharing and collective trust. Listen here.
Cloud Security Podcast: Closest alligator to the canoe: How transforming the SOC became P0 for Lloyds Bank: Matt Row, chief security officer, Lloyds Bank, explains the bank's digital transformation strategy, highlighting how it modernized its security operations center to achieve a 20x reduction in human-reviewed alerts. Listen here.
Defender’s Advantage: Human-machine teaming and applying AI to frontline threat intelligence workflows: Jake Nicastro, AI lead, Frontline Intelligence Operations, GTIG, details how his team is shifting from simple prompt engineering to more advanced agentic workflows, focusing on a model of human-machine teaming. 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.

Three lessons in accelerating foundation model upgrades

Thu, 16 Jul 2026 16:00:00 +0000

Have you run into problems migrating your products from one model to the next?

Upgrading to the latest AI models is rarely simple. For engineering teams, model updates whether migrating to an entirely new model or updating to a newer checkpoint within the same model family, like moving from an earlier Gemini version to Gemini 3.5 — often require a slow and costly process of testing, proving quality, and manually evaluating new responses. For most engineering teams, upgrading to a new model checkpoint means months of manual toil to verify performance. And the industry is moving at breakneck pace – since 2023, we’ve announced six major model evolutions, bringing us to Gemini 3.5 today.

Our team at Google Cloud, Applied ML, has a goal to deliver transformative infrastructure and services that benefit both Google and our customers globally. As part of that, our team built an agentic workflow that completes model upgrades in hours instead of months.

In this blog, we’ll show you our approach and three lessons you can apply to accelerate your own foundation model upgrades using Gemini Enterprise Agent Platform — our new, comprehensive platform to build, scale, govern, and optimize agents – and Google Antigravity, our primary solution for developers using AI for coding and agent orchestration.

Three lessons in building a flexible agent system

To support different team needs, we had to rethink traditional automation and learned three key lessons along the way:

Lesson 1: Start with hands-on discovery. First, our engineers worked closely with product teams on real migration problems. This hands-on work helped us identify complex requirements and build our first guidelines for prompt optimization.
Lesson 2: Beware the rigidity of traditional automation. We turned these guidelines into a standard, automated workflow. While this version gave us some quick wins, we soon found that traditional automation was too rigid to handle different data formats and unique edge cases.
Lesson 3: Pivot to a flexible agent architecture. The real progress came when we rebuilt the tool using a flexible agent. Instead of forcing teams into a rigid process, the agent adapted to specific project needs, helping analyze data and test prompts dynamically with a high degree of adaptability.

How our partner teams cut migration time while boosting quality

Our partner team, which manages video translation and dubbing services, had an interesting challenge: their workflow required rewriting translated text so that the spoken duration matched the original video's pacing exactly, without altering the meaning. Historically, this strict constraint required maintaining a fine-tuned model. Their goal was to migrate to the latest out-of-the-box foundation model, guided purely by prompt engineering.

Using this agentic framework, the team provided their ground-truth dataset and baseline prompt. The system autonomously hill-climbed the prompt quality, migrating the service away from the custom stack

Make your own migration workflow with Agent Platform and Google Antigravity

These learnings can be applied by any engineering team looking to accelerate their own model upgrades. If your organization is struggling to keep pace with new foundational models, replacing manual toil with intelligent automation requires treating migration as an agentic workflow.

To build your own automated migration pipeline, follow these steps:

Deploy Autoraters: Pivot from manual human review to model-based Autoraters to evaluate the quality of a new checkpoint at scale and in a fraction of the time.
Build an agentic loop: You can use the Agent Development Kit within Gemini Enterprise Agent Platform to create your agent.
Automate the orchestration: To make the process even easier, leverage Antigravity to automate the underlying coding and agent orchestration and add in features such as loss reporting or headroom reports.

By shifting away from a manual, line-by-line engineering task, organizations can reduce infrastructural tech debt and confidently keep pace with the frontier of AI.

_{This work is the result of collaboration across Google. We thank key contributors: Anthony Green, Chris Lamb, Chungyen Li, Connie Huang, Elaine Han, Elena Erbiceanu Tener, Eugene Ie, Francesca Ciacchella, Igor Karpov, Jeanie Jung, Jose Menendez, Kiam Choo, Lina Sanders-Self, Longfei Shen, Martin Nikoltchev, Mason Ng, Matt Mancini, Paul Zhou, Pedram Oskouie, Samuel Smith, Tom Lawrie, Ye Tian, Zhen Lin}

IDC: Why the right networking approach is foundational to agentic AI

Wed, 15 Jul 2026 16:00:00 +0000

Editor’s note: Today we hear from IDC on the results of its 2026 AI in Networking Special Report Survey exploring the enterprises' concerns about networking infrastructure to support the rise of agentic AI in their organizations. The survey was sponsored by Google Cloud.

Enterprises are moving quickly on AI pilots, but the move from pilot to production remains uneven. While AI models remain important, IDC research indicates that the pilot-to-production bottleneck is primarily infrastructure-centric, with core networking concerns emerging as one of the leading drivers of AI project delays and abandonment. In IDC's 2026 AI in Networking Special Report Survey:

32.6% of respondents cite security concerns: As AI workflows become more distributed and autonomous, enforcing consistent security and governance becomes more difficult.
26.8% of respondents cite challenges in automation: Manual operations and fragmented controls can slow deployment and make AI environments harder to scale.
24.7% of respondents cite staff time and talent restrictions: Limited skills and operational bandwidth can constrain an organization's ability to move AI initiatives into production.

Agentic AI specifically heightens these concerns by introducing more distributed and dynamic interactions across applications, services, APIs, tools, and data sources. In production environments, these interactions often span different agent frameworks, model providers, clouds, open-source tools, SaaS APIs, and internal applications, expanding both the operational scope and the security and governance surface area.

Networking for operational control, security, and governance at scale

Networking is the primary enabler of agentic interactions and plays a foundational role for intracloud and intercloud network- and services-layer connectivity, end-to-end security, and consistent governance. In agentic systems, networking increasingly extends into tighter service-centric controls that govern how distributed services identify one another, communicate, and exchange data securely. While AI workloads in general are increasing east-west traffic demands, agentic AI adds an additional layer of complexity by creating dynamic interactions that require tighter policy, visibility, and control closer to the application workflow.

From an infrastructure perspective, networking is much more than just a connectivity function. It is part of the infrastructure platform control plane that applies policy-based controls, supports observability, and helps maintain consistent security and governance across an AI agent's activity. This is significant because framework-level controls alone become insufficient in environments where agents and services span different runtimes, clouds, deployment models, and operating domains.

That is why an infrastructure-level approach becomes key. It does not replace application frameworks or orchestration environments, but it provides broader and more consistent policy implementation across a complex architectural landscape. As agentic AI becomes more autonomous and distributed, organizations need these controls built in as part of the infrastructure to reduce fragmented observability, inconsistent policy application, and unmanaged shadow agent activities. From a cloud infrastructure standpoint, this is where cloud network services become strategically important.

Balancing act: A platform vs. best-of-breed approach

Agentic AI systems are inherently fragmented because of underlying distributed workflows. Enterprises are already navigating a rapidly evolving landscape of business requirements, open-source components, emerging protocol standards, and new architecture patterns. In this context, choices between best-of-breed point solutions and platform-based approaches should be strategic rather than ideological.

Best-of-breed capabilities may be necessary to address specific technical requirements. But it is also true that point solutions introduced across a distributed agentic AI landscape can create inconsistent policies, operational complexity, and governance gaps. IDC research reflects this tension. In IDC’s 2026 AI in Networking Special Report Survey, organizations remained divided between platform and best-of-breed preferences for AI workloads; among respondents who favored platforms, the main reasons cited were stronger security (32.9%), reduced complexity (27.7%), and faster deployment (24.2%).

In IDC's view, a balance is important. Platforms can provide a consistent operational and policy foundation for AI deployments, but at the same time, they need to be modular and extensible to allow the inclusion of best-of-breed functionality as part of the platform toolset. The right platform for agentic AI should be open, flexible, and able to evolve. It should support integration with third-party and open-source tools, allow insertions of needed security and observability functions, and adapt without complete architectural rework.

This is a period of technology disruption. Businesses must meet their AI objectives while carefully managing dynamic agentic AI systems. In this environment, networking not only remains a connectivity piece of the AI infrastructure but becomes foundational to how organizations establish operational control, apply policy consistently, and maintain end-to-end trust across agentic workflows.

As agentic AI systems continue to evolve, the demands they place are unlikely to be addressed through best-of-breed point solutions alone. Operationalizing agentic AI at scale will require organizations to leverage the right networking approach, supported by infrastructure platforms that are open, flexible, and extensible, enabling a cohesive and adaptable security and governance framework.

Message from the sponsor
The autonomous and non-deterministic communications of agentic applications pose challenges for which the infrastructure and governance models of the cloud-native era are not prepared. In the agent-native era, an infrastructure-led approach is required to enable agentic applications at scale in production with effective governance and observability. An extensible platform based on open standards is critical in enabling the agentic journey today and through its maturity. Learn about the infrastructure imperatives and open standards that make a viable agentic infrastructure here.

Google named a Leader in the 2026 IDC MarketScape for Worldwide Foundation Model Software

Tue, 14 Jul 2026 18:00:00 +0000

For years, we’ve built with a clear priority: putting the practical needs of the enterprise first. Long before generative AI dominated the headlines, we were focused on building the global infrastructure, security frameworks, and data platforms that power the world's largest organizations. We’ve always believed that technology is only as good as its reliability, security, and predictability in production.

By anchoring our frontier research to this enterprise foundation, we can deliver models built specifically for business impact. We believe that approach is why Google has been named a Leader in the IDC MarketScape: Worldwide Foundation Model Software 2026 Vendor Assessment¹ and highlights our history of turning cutting-edge frontier research into secure, production-grade systems that developers can deploy at scale.

We believe Google’s position as a Leader highlights the exact momentum we are seeing in the market and validates the unique strength of our integrated, first-party AI stack.

By translating Google DeepMind’s continuous pipeline of fundamental research into production-grade business products, we unite our robust infrastructure and foundation model software to work together seamlessly as a single, unified solution.

Gemini Enterprise: A unified system for the agentic era

A great foundation model is only as valuable as an organization's ability to safely put it to work. In the enterprise, that value is realized when models are given the tools, memory, and agency to act as autonomous partners – moving from simple prompt-and-response text to dynamic agents that can execute complex business workflows.

Gemini Enterprise serves as this end-to-end system. It brings our most powerful developer capabilities and user-facing tools together into a single architecture, featuring the Gemini Enterprise app as the front door for everyday business teams to interact with AI, and the Gemini Enterprise Agent Platform for developers to orchestrate them behind the scenes.

Agent Platform abstracts away the underlying complexity of how technical teams build, scale, govern, and optimize these agents – whether they are handling customer-facing workflows or managing internal operations. Any agent engineered on the platform can be instantly surfaced in the Gemini Enterprise app, giving your workforce immediate access to secure, custom-built tools. Because rigorous governance, enterprise security, and cryptographic identity are baked into the foundation by default, organizations can stop worrying about managing technological risk and start focusing entirely on driving agent-led business outcomes.

Powered by Gemini

Our latest models are built specifically to orchestrate and execute complex, multi-step actions. At I/O this year, we kicked off the Gemini 3.5 series with the release of Gemini 3.5 Flash. It delivers intelligence on multiple dimensions at speeds you have come to expect from the Flash series. It’s ideal for tackling long-horizon agentic tasks. Google DeepMind engineered these models from the ground up using our purpose-built AI infrastructure. This unique co-design of the model and hardware allows us to train deeper reasoning capabilities faster and more efficiently with every new generation.

Developers can build agents using Gemini 3.5 Flash on the Gemini Enterprise Agent Platform, or use it in your projects in Google AI Studio and Antigravity.

Business users can use Gemini 3.5 Flash in the Gemini Enterprise app to help discover, create, and use the best of Google AI in their workflows starting today.

Get started

Download the IDC MarketScape: Worldwide Foundation Model Software 2026 Vendor Assessment excerpt to learn why organizations are choosing Google Cloud.

Explore Gemini Enterprise today, or speak to your Google Cloud account representative to schedule a hands-on technical workshop.

^{1. IDC MarketScape: Worldwide Foundation Model Software 2026 Vendor Assessment, Doc #US54427726, July 2026}^{IDC MarketScape vendor analysis model is designed to provide an overview of the competitive fitness of technology and suppliers in a given market. The research methodology utilizes a rigorous scoring methodology based on both qualitative and quantitative criteria that results in a single graphical illustration of each supplier’s position within a given market. The Capabilities score measures supplier product, go-to-market and business execution in the short-term. The Strategy score measures alignment of supplier strategies with customer requirements in a 3-5-year timeframe. Supplier market share is represented by the size of the icons.}

Claude at scale on Google Cloud: Frontier AI, built for enterprise production

Tue, 14 Jul 2026 16:00:00 +0000

Running frontier AI in production is demanding — accelerators to manage, latency to hold steady across continents, regulated data to keep in-region, and long-context requests to serve reliably. Claude on Google Cloud is built for exactly this.

Like Monet and water lilies, frontier models and the enterprise platforms are often better together. In our case, Claude brings the reasoning, and Google Cloud brings the managed infrastructure, global reach, and compliance posture that enterprises already run on. Calling Claude becomes operationally identical to calling any other Google Cloud service — same Identity and Access Management (IAM), same VPC Service controls, same observability — so teams are able to spend their time building features instead of running inference infrastructure.

This post walks through what Claude on Google Cloud delivers in production across four areas:

Managed infrastructure that gives engineers their time back
Global endpoints that hold latency low, and uptime high for a worldwide user base
Security and data-sovereignty controls inherited straight from Google Cloud
Serving-layer features that keep cost and performance optimized at scale.

Managed infrastructure that frees engineering time

Claude on Google Cloud runs on fully managed infrastructure, so enterprise teams ship features instead of building clusters. Compute provisioning, auto-scaling logic, load balancing, and failover at frontier-model scale are handled by the platform — work that would otherwise occupy multiple teams full-time.

Claude is available through Agent Platform's Model Garden as a Model-as-a-Service offering, ready to use over standard REST / JSON over HTTP/1.1 or HTTP/2 endpoints. Invoking Claude is operationally identical to invoking any other Google Cloud service: the same IAM policies, the same VPC controls, and the same observability stack via Cloud Logging and Cloud Monitoring.

Serving Claude takes a few lines of Python using the AnthropicVertex client:

code_block: <ListValue: [StructValue([('code', 'from anthropic import AnthropicVertex\r\n\r\nclient = AnthropicVertex(\r\n project_id="your-project-id",\r\n region="us"\r\n)\r\n\r\nmessage = client.messages.create(\r\n model="claude-opus-4-8",\r\n max_tokens=1024,\r\n messages=[{"role": "user", "content": "Analyze this system architecture."}]\r\n)'), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7f8142ce5e20>)])]>

The same AnthropicVertex client handles prompt caching, tool use, structured outputs, streaming, and adaptive thinking; for batch inference, use Vertex AI Batch Prediction. Authentication uses Application Default Credentials; requests automatically inherit your project's IAM and VPC configuration.

Global reach with consistent latency and built-in failover

Serving a worldwide user base from a single endpoint produces high tail latency and a single point of failure. Most enterprises can't replicate inference infrastructure across continents while keeping performance consistent.

Agent Platform exposes three endpoint types for Claude, each solving a different production requirement:

Global endpoints route requests to a region with available AI compute capacity. For example, if us-central1 is capacity-constrained, traffic redirects to europe-west1 or another region with available capacity. That’s automatic failover and geographic load balancing without application-side routing logic. Global endpoints are ideal for maximum availability and lowest cost.
Regional endpoints like us-east5 or europe-west1 keep prompts, completions, and intermediate state inside a specific geographical boundary, making it ideal for low latency and data-residency requirements.
Multi-region endpoints give U.S. or EU data residency without single-region dependency. They dynamically route across regional endpoints providing built-in resilience against regional outages and capacity constraints.

The diagram below shows how applications reach Claude through these endpoint types, and how the Agent Platform serving layer routes traffic to the Compute AI clusters across regions:

Serving Claude Models From Regional & Global Endpoints

Serving Claude Models From Multi-Region Endpoints

Serving Claude Models From Regional Endpoints

Enterprise security and data sovereignty built in

Regulated workloads — financial services, healthcare, and government — get enterprise-grade security and data sovereignty without trading compliance for convenience, and without re-engineering the hardest layer to control: inference, where prompts, completions, and intermediate state all flow through the serving stack.

Claude on Agent Platform inherits Google Cloud's full security posture. FedRAMP High and HIPAA compliance enable deployment in government, healthcare, and financial services environments. VPC Service Controls let organizations define a perimeter around Agent Platform resources, preventing data exfiltration. IAM-native access control governs Claude endpoints with the same roles and policies that protect every other Google Cloud resource — no separate API keys to manage or rotate. Cloud Logging and Cloud Monitoring provide near real-time visibility into token usage, error rates, latency, and quota consumption.

Combined with the regional and multi-region endpoints above, this gives regulated customers a path to running frontier AI in production without re-auditing their compliance posture.

Optimized for cost and performance at scale

In production, cost and performance drive every architectural decision. Getting both right requires capabilities from two layers: Claude's native model features, and Google Cloud's serving infrastructure. Agent Platform supports both, so teams can optimize across the stack without managing them separately.

Claude-native capabilities, fully supported on Agent Platform

These features are built into Claude and available on Agent Platform without any additional configuration:

Prompt caching stores and reuses shared prefixes — long system prompts, legal documents, codebases — reducing request latency by up to 80% and cost by up to 90%.
Streaming responses over server-sent events deliver tokens as they are generated, critical for chat interfaces and coding assistants where perceived latency matters.
Extended and adaptive thinking lets Claude dynamically determine when and how much to reason through complex, multi-step problems — and allows users to dial the thinking effort directly, for example to control cost. Optimized for use cases like advanced code generation, mathematical reasoning, and multi-document analysis.
Extended context windows up to 1M tokens (for Claude Opus 4.6,Sonnet 4.6 and newer models) enable long-document analysis, large codebase reasoning, and multi-turn conversations at depth.

Google Cloud serving infrastructure

Agent Platform adds its own serving-layer capabilities on top of Claude's native features:

Batch prediction handles large-scale offline workloads — document classification, content moderation, bulk summarization — asynchronously at lower priority and reduced cost.
Provisioned throughput reserves dedicated inference capacity for mission-critical workloads, isolating them from public traffic and ensuring predictable performance during peak demand.
Memory management and scheduling for long-context requests is handled at the infrastructure layer,.

Together, these two layers give teams the full range of optimization levers — from model-level efficiency to infrastructure-level capacity control — on a single, unified platform.

From inference to agents

The same infrastructure that serves Claude inference powers the agent layer of Agent Platform on Google Cloud. The build-and-register flow has three steps:

Build with Claude. Claude is well-suited as an orchestration backbone — its extended context window, native tool use, and adaptive thinking make it effective at planning multi-step tasks and delegating to sub-agents. Pick Claude Opus, Sonnet, or Haiku from the Model Garden, then build with the Agent Development Kit (ADK) — code-first in Python, Go, Java, or TypeScript — deploy to Agent Runtime, Cloud Run or Google Kubernetes Engine.
Deploy the Agent to a Runtime. Depending on your use case, select Agent Runtime, Google Kubernetes Engine or GKE Agent Sandbox to run your deployed agents.
Interoperate over A2A. The Agent2Agent protocol runs at 150+ organizations, letting a registered Claude-powered agent delegate tasks to agents from SaaS and other service providers.

The result: a planning agent built on Claude can orchestrate sub-tasks across the broader agent ecosystem, under unified IAM, fully auditable, on the same infrastructure that serves the underlying inference.

Start building

Open the Agent Platform console, enable Claude in the Model Garden, and make your first API call with the AnthropicVertex SDK. Add prompt caching, provisioned throughput, and other features as your workload demands. When you're ready to go agentic, learn more about Claude on Agent Platform.

Reach out to your Google Cloud sales representative to discuss bringing Claude into your production environment at scale.

Securing the AI supply chain on GKE: Introducing k8s-aibom for automated AI BOMs

Mon, 13 Jul 2026 16:00:00 +0000

How should your security team manage shadow AI? Workloads deployed by developers without formal registration can often evade traditional security scanners, because organizations are reluctant to slow down development and compromise stability by demanding privileged Daemonsets, kernel-level access, and manual pod-spec edits.

To break this deadlock, today we are open-sourcing k8s-aibom. This lightweight, unprivileged Kubernetes controller continuously monitors the cluster API and container environments to automatically detect running AI runtimes (like vLLM and Triton) and generate standard CycloneDX Machine Learning Bill of Materials (ML-BOMs).

By providing automated, audit-grade visibility directly from runtime execution — regardless of whether the workload was formally registered — k8s-aibom can help teams safely move AI projects from pilot to production without developer integration friction.

The architecture of zero friction

k8s-aibom is designed from the ground up to respect both the CISO mandate for total visibility and the SRE mandate for cluster stability. It deploys as a single, unprivileged Deployment in the k8s-aibom-system namespace. It involves zero developer friction — no sidecars, no eBPF kernel modules, no privileged DaemonSets, and no modifications to existing developer pod specifications.

k8s-aibom watches for AI workloads and produces BOMs.

The discovery pipeline executes through four clear stages:

Scrape cluster workloads: The controller continuously monitors KServe resources, Deployments, StatefulSets, DaemonSets, and Jobs across the cluster.
Identify AI stacks: Advanced pattern matching inspects container images, environment variables, and command-line arguments to detect serving runtimes (vLLM, Triton Inference Server, TGI, Ollama), autonomous agent frameworks (LangChain, AutoGen, CrewAI), vector databases and RAG stores (Milvus, Qdrant, pgvector), as well as distributed training jobs and evaluation harnesses.
Generate standard manifests: The controller compiles the discovered artifacts into formal OWASP CycloneDX 1.6 Machine Learning Bill of Materials (ML-BOM) documents.
Export to sinks: The controller attaches the resulting ML-BOM directly to the custom resource status (status.bomDocument) of an in-cluster AIBOM Custom Resource (CR) and routes it to optional external sinks, including Google Cloud Storage buckets and external webhook endpoints.

Application teams do not need to modify their pod specifications, inject sidecar containers, or alter their continuous integration and continuous delivery (CI/CD) pipelines. Furthermore, k8s-aibom treats the Kubernetes cluster state as a pure functional input: Identical cluster inputs produce byte-identical ML-BOM documents. This deterministic property makes k8s-aibom an ideal fit for GitOps workflows, enabling site-reliability engineers (SREs) to perform exact diffs and trigger precise change-detection alerts when AI dependencies drift.

Where existing AIBOM tooling falls short

Many AI BOM solutions offer build-time scanners producing BOMs from artifacts at rest. These tools help you track the code that was intended to be deployed.

Commercial AI security platforms extend the picture with cloud-native posture management, but typically through external scanning shaped around vendor-specific data models. Few, if any, of these tools help compliance reviewers, security operations (SecOps) teams, and platform engineers understand what is running right now, what is it connected to, and how can we verify those assertions.

We purpose-built k8s-aibom to bridge that gap. It produces BOMs from live cluster observation rather than artifact scanning, emits standards-conformant CycloneDX 1.6 ML-BOMs that integrate with the broader OWASP and Open Source Security Foundation (OpenSSF) supply-chain ecosystem rather than vendor-proprietary formats, and runs as an unprivileged controller on any conformant Kubernetes cluster — making it complementary to existing build-time and posture-management tooling rather than a replacement for either.

The Confidence Model: Separating intent from inference

For compliance auditors and SecOps engineers, raw telemetry is often noise. Standard monitoring tools indicate that a container is running, but can’t prove whether an AI model was explicitly configured by a platform engineer or dynamically pulled by an autonomous script at runtime. k8s-aibom solves this ambiguity through its deterministic Confidence Model, categorizing discovered assets into distinct tiers:

Declared: Explicitly defined by the customer or developer in the workload configuration (For example, explicitly passed container arguments such as --model meta-llama/Llama-2-7b.) A “declared” confidence detection represents clear human intent.
Inferred: Derived autonomously by the controller's pattern-matching engine through deep inspection of container images, environment variables, and execution profiles. (For example, identifying ^vllm/.* container signatures.)
Unresolved: Applied to workloads where an active AI presence is detected, but exact model parameters, weights, and versions can’t be deterministically established. An “unresolved” confidence detection immediately flags the workload for targeted security review.

This structured taxonomy allows compliance reviewers to instantly separate explicit engineering intent from machine inference, establishing an unassailable chain of trust during audits.

Immutability and least privilege: Building an audit-grade security model

Auditors remain deeply skeptical of standard observability telemetry because logs and metrics can be modified, dropped, and tampered with by compromised nodes or elevated administrators. k8s-aibom establishes an audit-grade evidence trail built on strict least-privilege isolation and data immutability.

The controller operates under a dedicated Kubernetes service account bound to a minimal Identity and Access Management (IAM) Workload Identity. It acts as the sole identity authorized to write BOM records to external storage sinks, requiring only roles/storage.objectCreator permissions.

To satisfy the most stringent audit and evidentiary standards, the Google Cloud Storage external sink implementation enforces DoesNotExist preconditions on object creation. Once an ML-BOM is written to the Cloud Storage bucket, the object becomes cryptographically immutable.

It can’t be silently overwritten, modified, or retroactively tampered with by compromised cluster actors or rogue workloads. SecOps teams gain absolute assurance that the historical audit log presented to regulators represents an unalterable record of cluster execution.

Accelerating governance readiness: Mapping to global regulatory frameworks

By automating the generation of standardized CycloneDX 1.6 ML-BOMs, k8s-aibom directly bridges the gap between low-level Kubernetes runtime state and high-level governance frameworks. It unblocks stalled GKE AI deployments by providing the foundational empirical data essential to major global standards:

EU AI Act: Designed to help organizations align with Article 12 (automated logging and record-keeping for continuous traceability) and Article 50 (transparency obligations for AI systems). By automatically cataloging serving runtimes and agent stacks, the tool helps simplify the gathering of technical evidence that may be needed during compliance audits.
NIST AI Risk Management Framework (AI RMF): Provides continuous, empirical asset visibility that can help support the Govern, Map, Measure, and Manage functions, helping shift compliance workflows from purely manual checks toward more automated asset inventory tracking.
ISO/IEC 42001:Supports compliance efforts for AI management system asset discovery and tracking, reducing the reliance on manual spreadsheets or periodic snapshot audits for inventory validation.

Getting started

It’s rare that a technical solution like k8s-aibom can help mitigate the multi-faceted problem of shadow AI, impacting CISOs, governance, risk, and compliance teams, SecOps teams, platform engineers, and developers.

To learn more by inspecting the controller, review the CRD definitions, and contribute to the open-source k8s-aibom project, please visit the k8s-aibom GitHub Repository.

Frontier and Center: Who evaluates the evaluations?

Fri, 10 Jul 2026 16:00:00 +0000

Editor’s note: Some of the most interesting questions in AI are being asked by information theoreticians, around how to provide context to an emerging class of AI agents. A few weeks ago, we waded into those waters with a blog about the Open Knowledge Format, a specification that formalizes the LLM-wiki pattern into a portable, interoperable format to represent the metadata, context, and curated knowledge that modern AI systems need to operate. That blog generated a ton of interest, so we’ve decided to bring you more of the same, as part of our new “Frontier and Center” series. Today, we hear from two members of Google Data Cloud’s frontier AI team on the recurring challenge of how to systematically evaluate whether or not an agent is able to answer questions effectively based on its context. Read on for more, and watch this space for more blogs from this team.

A passing grade is the least interesting thing an exam can tell you. It says the student cleared the bar; leaving you entirely in the dark about how narrow their failures were, how effortless their passes were, or what to teach next. Yet this is exactly how we evaluate AI agents. We run a fixed benchmark, calculate a score, and declare progress. In doing so, we are handing our agents a pass/fail exam when what we actually need is a map of the agent’s capabilities: a picture of the terrain that shows exactly where capability falls off, and by how much.

For data agents, this map matters a lot for data discovery in search and retrieval — the unglamorous first step where an agent, handed a vague human question and a warehouse or data lake of thousands of tables and files, has to find the right datasets before it can reason over anything. Discovery is a "needle in a haystack" problem. Real users phrase their questions imperfectly, and inferring what datasets to retrieve presents a real challenge to agents. So the interesting question in evaluations is never "can the agent pass?" It is "how vague can the question get before the agent breaks?" An exam cannot easily answer that, but a map can.

Today, we share an approach rooted in information theory that we’ve been leveraging to add detail and nuance, i.e., fidelity, to benchmarks, so we can better understand agents’ performance as a part of their evaluations. Along the way, the added fidelity exposed some deeper issues with the quality of emergent evaluation cases themselves.

Difficulty, measured

When it comes to retrieval, evaluation cases are often stratified into tiers of difficulty. This can happen organically, e.g., pervasive and enduring failure scenarios are deemed difficult. Or it can be from labels applied by humans or machines categorizing some questions as "easy" or "hard" for an agent to answer correctly, e.g., based on the context provided in the query. While this kind of sentiment-based labeling is not the only way to label test cases, it’s frequently used despite its imperfections, such as being challenging to reproduce.

Despite being an industry staple, the approach of assessing every evaluation case by hand is unrealistic at scale. What we need is a rigorous approach that can modulate the difficulty of evaluation cases. We’re iterating on a meta-benchmark we call Discovery Bench: a framework that modulates an evaluation case by generating “easy” and “hard” variations of every case. This allows us to audit how close or how far an agent is from succeeding in those cases.

The lever for modulating the difficulty of an input query comes via a tried-and-trusted concept that’s present across information theory and machine learning: surprisal, or the likelihood of an output given a set of inputs. In our case, a query’s surprisal represents the uncertainty that remains about the correct dataset given the query.

The thinking behind our approach is simple: A term or a phrase in an evaluation query has high informative power when it sharply distinguishes the target from everything else in the corpus. Therefore, we can adjust the difficulty of evaluation cases by adding or removing terms with varying levels of informative power.

Let’s work through a real example from KramaBench, a publicly available benchmark. One of KramaBench’s datasets has information about orbiting satellites, and the example query from the suite includes the following text: "…the total count of satellite major altitude changes for satellite 48445 during 2024 using TLE history."

The token "TLE" is sharply distinguishing; it points almost uniquely at the TLE_____48445 table from the dataset. Strip it, and the query degrades to "the count of satellite altitudes for satellite 48445," whose vague phrasing now matches density tables, precise-orbit files, and decay logs alike. Surprisal makes this quantitative: rare, pointed terms carry more bits than common ones.

The remaining surprisal of a query is how much uncertainty is left about its answer. As surprisal approaches zero, the query has become specific enough to pinpoint exactly one dataset.

The heart of the idea behind Discovery Bench is this refinement loop, which we call iterative surprisal-based query refinement, or iSQR, which generates cases with higher or lower informative power to test where an agent can start successfully answering the query:

Figure 2: The iSQR refinement loop.

The crux is being able to control the challenge embedded into the evaluation case by making adjustments: Instead of one fixed phrasing per question, we generate the same question at three levels of calibrated ambiguity [high, medium, low], with each grounded in bits (not subjective opinion). We can even justify, term by term, why a word was added or removed. Difficulty stops being a property that is attributed by sentiment or classification, and becomes one we engineer.

The cliff you couldn't see

Here is what Discovery Bench’s difficulty dial reveals — and what a single-phrasing benchmark structurally cannot.

We have an F1 agent that's built for recall (on Gemini 3.1 Pro). Running it against KramaBench and across the full sweep of ambiguity levels traces a curve: 0.34 at high ambiguity, 0.76 at neutral, 0.81 at medium, 0.78 at low.

Figure 3: F1 swept across ambiguity — the dot versus the curve.

Two findings fall out immediately (and neither were visible to a conventional eval).

First, the cliffs. This query scores a perfect F1 = 1.00 at neutral phrasing — and 0.00 at high ambiguity. It is the satellite-48445 case from above: drop the distinguishing token "TLE" and the agent loses the table entirely. Same query, same agent, same ground truth; one notch vaguer and it falls off a cliff. A static benchmark tests the neutral phrasing, stamps "solved," and reports flat ground where there is a precipice. Pass/fail was particularly misleading in that it did not just miss the cliff, but it told us the terrain was level.

Second, the sweet spot. For Discovery Agent, medium ambiguity beat neutral, and low ambiguity sometimes underperformed it. More specificity is not monotonically better for the system being evaluated; there is an optimal amount of steering. That is a graded, actionable signal. This is the "how close, how hard" texture we were missing from a scalar. It tells you where to hill-climb, or improve, the agent: in our case, straight at concrete failure modes like time-sharded tables (precision collapsing to ~8% as the agent over-retrieves 21 near-identical shards for a two-table answer) and context blow-up (F1 dropping from 0.75 to 0.32 once a query triggers long search chains). The map did not just say that the agent failed, but it said where, and why. Note that our hypothesis that less ambiguity and more context (via steering terms) should improve retrieval generally holds true, but for the specific Discovery Agent being exercised, the idiosyncratic “sweet spot” meaningfully highlighted trade-offs in its implementation.

We're not alone

The field is converging on meta-benchmarking and exerting greater control of how we challenge and evaluate our agents. A growing body of work uses item response theory, the latent-ability model behind standardized testing, to treat difficulty as a measured quantity rather than a label: tinyBenchmarks and metabench show that a handful of informative items reproduce a model's full score, and PSN-IRT turns the same lens on benchmark quality itself. Others audit the ground truth directly: MMLU-Redux found that 6.49% of Massive Multitask Language Understanding (MMLU) questions are mislabeled, and Platinum Benchmarks re-cleaned ten datasets to minimize both label errors and ambiguity — the same two axes we sweep for. And ambiguity is increasingly treated as intrinsic rather than noise: AmbigQA showed that a large fraction of real questions admit multiple readings, and later work finds that apparent hallucinations often stem from query ambiguity rather than model failure. What we have not seen elsewhere is the combination: information-theoretic ambiguity sweeping applied as a meta-benchmark over live enterprise data.

A benchmark we trusted turned out to be broken

We built our first evaluation on kramabench-astronomy, a benchmark established in the field, and one which other teams had already leaned on for their own evals. Teams derived benchmarks from this dataset, and we hypothesized subtle issues may have been introduced over time. When we actually read the benchmarks used by teams, with Gemini's help, we found it was wrong in meaningful ways: ground-truth tables that did not answer their query, a question whose 124 sharded tables exceeded what some teams’ retrieval APIs could even return, months specified where exact dates were required. Quietly broken ground truth means quietly wrong conclusions not just for us, but for every prior analysis built on it.

This is the generalized crux of the matter: an evaluation is itself an artifact that can be defective, and almost nobody evaluates it. We instrument the agent and trust the ruler, but where do we validate that the measuring stick makes sense?

When two maps disagree

Now the recursive turn: If difficulty is something we generate, then we need to evaluate the generator itself; we should not trust it blindly either.

So we built the same ambiguity sweep two ways: steering terms from a pure-LLM guess, versus terms grounded in TF-IDF surprisal. The two disagreed violently. At high ambiguity, the LLM-built sweep scored the agent at F1 ≈ 0.34; the grounded sweep, ≈ 0.85. One of these maps is badly distorted. The grounded one, predictably, is the more robust: surprisal gives it a footing the free-running LLM lacks.

This is "evaluate your evals," made concrete. The information-theoretic lens does not only grade the agent along a continuous axis; it grades the benchmark's own construction, and adjudicates between the two.

Evaluate your evals

We have spent years optimizing agents against rulers we never measured. The bitter irony is that better models make this worse: as agents clear coarse benchmarks, the score saturates near the top and the exam loses its ability to highlight where the agent can be improved.

So the call to action is uncomfortable and overdue: evaluate your evals. Read your ground truth. Treat difficulty as a measured quantity, not a label: sweep it, plot it, find the bit-width where your system breaks. Ask not just "did it pass?" but "how close was the miss, how hard was the pass, and would a slightly vaguer question have sent it off a cliff?" Build evaluations that produce signals; not just verdicts.

There is a genuine tension to sit with here. Difficulty-as-entropy is only as reliable as the model that estimates the entropy. There's a risk that if we push too hard on a measurable proxy, we optimize the ruler instead of the agent. That is not a reason to retreat to pass/fail; it is a reason to keep the evaluator under the same scrutiny as what it is evaluating. The moment we stop asking who evaluates the evaluators is the moment our maps stop being useful again.

^{1. Maia Polo, F. et al. tinyBenchmarks: Evaluating LLMs with Fewer Examples. ICML 2024. arxiv.org/abs/2402.14992}^{2. Kipnis, A. et al. metabench: A Sparse Benchmark of Reasoning and Knowledge in Large Language Models. ICLR 2025. arxiv.org/abs/2407.12844}^{3. Lost in Benchmarks? Rethinking Large Language Model Benchmarking with Item Response Theory (PSN-IRT). AAAI 2026. arxiv.org/abs/2505.15055}^{4. Gema, A. P. et al. Are We Done with MMLU? (MMLU-Redux). 2024. arxiv.org/abs/2406.04127}^{5. Vendrow, J. et al. Do Large Language Model Benchmarks Test Reliability? (Platinum Benchmarks). 2025. arxiv.org/abs/2502.03461}^{6. White, C., Dooley, S. et al. LiveBench: A Challenging, Contamination-Limited LLM Benchmark. 2024. arxiv.org/abs/2406.19314}^{7. Min, S. et al. AmbigQA: Answering Ambiguous Open-domain Questions. EMNLP 2020. aclanthology.org/2020.emnlp-main.466}^{8. Lai, E., Vitagliano, G. et al. KramaBench: A Benchmark for AI Systems on Data-to-Insight Pipelines over Data Lakes. 2025. arxiv.org/abs/2506.06541}

Solve harder problems with AlphaEvolve, now available to everyone on Google Cloud

Thu, 09 Jul 2026 16:00:00 +0000

Many of the most challenging and valuable problems in the world are related to optimization. Now, AI is making these problems tractable. If you've ever tried to design a microchip, plan a delivery network, or optimize a training architecture for a large AI model, you know how hard it is to find the most optimized code. Traditional coding methods often cannot explore all the possible algorithms and implementations because the search space is simply too vast. To help, we introduced AlphaEvolve last year in private preview — an agent to help you design better algorithms on Google Cloud.

What’s new: Today, AlphaEvolve is generally available (GA) on Google Cloud. AlphaEvolve is a code optimization and discovery agent built on top of Gemini that helps solve the hardest algorithmic problems and achieve breakthroughs for your business and research. It has been tested in diverse domains like logistics, semiconductors, genomics, high performance computing, and financial services during our early access program. It systematically explores the search space to find solutions optimized for your problem.

Deploying AlphaEvolve within your environment follows a structured four-step process designed to move from initial problem definition to fully optimized production code:

Define: Provide a baseline seed algorithm and problem definition, together with background knowledge that provides context about the problem you want to solve.
Measure: Establish a scoring function to objectively score candidate programs on one or more metrics important for your problems such as correctness, performance, and operational constraints.
Optimize: Use AlphaEvolve’s agentic harness to generate optimized code, explicitly optimized against the metrics in the scoring function established in the measure step.
Apply: Deploy the resulting, highly optimized algorithm directly into your production workloads and infrastructure.

In this post, we’ll share how organizations are already seeing impact with AlphaEvolve and how you can get started.

How organizations are using AlphaEvolve

AlphaEvolve has grown from a research project into a key tool we use at Google. Now, some of the world’s most innovative organizations are using it to solve their algorithmic problems, too.

BASF: Building a digital twin to optimize global supply chains
"We had several attempts to build a digital twin for our complex supply network using deterministic models, and all of them failed. By using AlphaEvolve, we can now not 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." — Dr. Goetz Krabbe, Vice President for Global Supply Chain, BASF.

Visit the blog to read more how BASF used AlphaEvolve to improve their existing planning and forecasting models by over 80%.

Coolblue: Optimizing e-commerce demand forecasting
“Coolblue data scientists used AlphaEvolve to directly optimize their 28-day demand forecasting pipeline, focusing on automated feature engineering, target preprocessing, and model selection. In just a few (200) iterations, AlphaEvolve improved our production forecast (by reducing WMAPE over the existing solution) by over 5%. These gains were achieved through improved feature engineering, an ensemble of different regression models, and better target preprocessing proposed and validated by AlphaEvolve. To ensure sufficient stock availability, it is crucial that the demand forecast is accurate for both the short term (the first 7 days) and the longer horizon (the full 28 days). AlphaEvolve achieved this by using an evaluation metric that combines both periods, along with a strict penalty for under forecasting. AlphaEvolve has proven its ability to significantly improve bulk purchasing decisions and help us maintain optimal stock levels for the weeks ahead.” — Cas Ruger, Data Scientist at Coolblue.

FM Logistic: Optimizing warehouse routing
"Through our partnership with Google Cloud and the implementation of AlphaEvolve and Gemini, we further optimized our routing approach for fast-moving operations. The 10.4% improvement was achieved on top of an already highly optimized baseline, where further gains are typically hard to come by. This translates directly to faster fulfillment, improved working conditions for our teams, and reduced wear on our fleet." — Rodolphe Bey, Group CIO at FM Logistic.

Visit the blog and website to read more about how FM Logistic used AlphaEvolve to improve warehouse routing by 10.4%, saving over 15,000 km in staff travel.

Infineon: Optimizing chip design
"Our initial experiments with AlphaEvolve have been very positive, demonstrating its potential to transform the chip design lifecycle. We see a clear potential for it to contribute to multiple stages of development, including areas like Surrogate modelling." — Michael Kollig, CIO, Infineon.

JetBrains: Accelerating IDE performance
"AlphaEvolve can change how we approach complex performance work. It turns optimizations that were once too time-consuming to explore into candidates we can test routinely. Engineers still own the benchmark, review, and release decision. The search space is what gets smaller." — Dmitrii Batkovich, Director of Engineering, JetBrains.

Visit the blog to read more about how Jetbrains used AlphaEvolve to improve their IDE performance by over 15-20%.

Kinaxis: Improving optimization and forecasting systems
"Kinaxis researchers have used AlphaEvolve to materially improve both the speed and quality of highly mature forecasting and optimization algorithms. In early testing, we achieved improvements of more than 22% in key forecasting accuracy metrics while reducing runtime by over 90% on benchmark datasets. As supply chains become increasingly complex and unpredictable, AlphaEvolve has the potential to help the world's largest organizations make faster, more informed decisions and adapt with greater confidence." — Gelu Ticala, Chief Technology Officer, Kinaxis

Visit the blog to read more about how Kinaxis used AlphaEvolve to achieve significant gains across their forecasting and runtime metrics.

Klarna: Doubling throughput while improving model quality
"Klarna applied AlphaEvolve to one of their largest ML training pipelines and doubled throughput while improving model quality, all under the strict reproducibility constraints of regulated financial services. Over three weeks, the system explored nearly 6,000 candidate programs, discovering deep architectural rewrites no engineer would have tried." — Klarna engineering team.

Visit the blog to read more about how Klarna used AlphaEvolve to double Training Speed and improve performance for their foundational models.

Kuro Games: Server-side Optimization
"At Kuro Games, our guiding principle is that AI should not just make our work faster — it should make our work better. AlphaEvolve is a real-world validation of that principle. We applied it to a complex backend optimization challenge and saw substantial performance gains in specific server-side workloads. AlphaEvolve handles the kind of optimization work machines do best, so our engineers can focus on what only people can do: crafting great games." — Lin Chenchen Chief Technology Officer, Kuro Games

Oak Ridge National Laboratory: GPU kernel generation for exascale computing
Under Google DeepMind’s Genesis Mission partnership with the Department of Energy to provide early-access to our AI for science tools.

“Oak Ridge National Laboratory (ORNL) recently partnered with Google to deploy AlphaEvolve on Frontier, the world’s first exascale supercomputer. The research team built a closed-loop evaluation architecture that bridges cloud-based large language model code generation with Frontier’s execution environment. The designed system optimizes mixed-precision GPU kernels—which requires complex, coupled decisions about memory, data layout, and hardware synchronization — by iteratively generating, compiling, running, and validating candidate programs, directly on the supercomputer's AMD GPUs. This executable search framework evaluates each proposed structural optimization against numerical accuracy rules.

“Our collaboration with Google's AlphaEvolve team gave us an early look at how evolutionary programming can be combined with leadership-class supercomputing. By running AlphaEvolve on Frontier, we explored a large number of optimization candidates in parallel, including novel implementation variants that helped us explore parts of the design space we might not have reached through manual optimization alone. This is an encouraging first step toward applying AI-assisted optimization to increasingly complex scientific software." — Oscar Hernandez Mendoza, PhD, Senior Computer Scientist, ORNL

Old Dominion University: Modeling biological aging mortality rates
"The Qin Lab at Old Dominion University used AlphaEvolve to search the space of Python programs that model biological aging mortality rates, a problem in computational biogerontology where the governing equations span multiple empirical laws. Utilizing an HPC cluster in Google Cloud as a part of the ODU MonarchSphere initiative, AlphaEvolve – across approximately 500 evaluations – independently rediscovered the Kannisto logistic mortality model (a published result from the 1990s biogerontology literature) with no prior knowledge of that literature, improved the Emergent Aging Model composite fitness score by 19% through heterogeneous decay rate distributions, and demonstrated near-perfect Strehler-Mildvan correlation (0.949) via scale-free network topology with Laplacian spectral aging across approximately 500 evaluations. The central finding is that structurally diverse models all converge on the same empirical aging laws, providing evidence that Gompertz, Strehler-Mildvan, and Kannisto regularities are robust attractors of biological systems. The team plans to extend this work to multi-species datasets and to connect the evolved program structures to testable biological mechanisms.” — Dr. Hong Qin,Department of Computer Science, Old Dominion University

PacBio: Scaling accuracy and lowering costs in genomics
"The solution the Google team discovered using AlphaEvolve unlocks meaningfully higher accuracy rates for our sequencing instruments. For researchers, this higher-quality data might enable the discovery of previously hidden disease-causing mutations." — Aaron Wenger Senior Director, PacBio.

Visit the blog to read more about how Pacbio used AlphaEvolve to improve DeepConsensus — a model developed by Google Research for correcting DNA sequencing errors — achieving a 30% reduction in variant detection errors.

Pebble: Optimizing serving performance on GPUs
"Optimizing inference serving is an incredibly challenging problem because it is a multi-dimensional system design challenge that shifts dynamically between memory, compute, and hardware orchestration constraints. NVIDIA's AI Configurator latency model was severely bottlenecked by a single, static 0.8 empirical correction factor that applied uniformly to all workloads, and did not model FP8-vs-BF16 efficiency divergence, causing recommended configurations to drift away from the optimum. AlphaEvolve solved this by autonomously discovering GPU performance modeling formulations directly from our training prior. This Gemini-powered evolutionary approach drastically cut our model errors by more than delivering a 56% relative error reduction. We are excited to integrate this smoother, learned efficiency function and leverage AlphaEvolve to continuously map emerging hardware specifications without manual tuning." — Keval Shah Head of AI, Pebble

Qbraid: Advancing quantum computing
"AlphaEvolve delivered a result on top of an encoding family we had already spent years refining. It searched a design space far too large to comb through by hand and handed back something we could read, verify, and understand. Systems like AlphaEvolve will meaningfully accelerate progress toward useful quantum computing." — Kenny Heitritter, Vice President of Research and Development at qBraid.

Visit the blog and paper to read more about how Qbraid used AlphaEvolve to find significantly more error efficient error-correcting codes for quantum chemistry.

Schrödinger: Shortening cycles for molecular simulations for drug discovery
"AlphaEvolve allows us to explore larger chemical spaces faster and more efficiently than ever before. Faster MLFF inference carries real business impact, shortening R&D cycles in drug discovery, catalyst design, and materials development, and enabling companies to screen molecular candidates in days rather than months." — Gabriel Marques, ML Tech Lead, Schrödinger.

Visit the blog to read more about how Schröedinger used AlphaEvolve to quadruple the speed of molecular discovery.

Substrate: Accelerating runtime speed for semiconductor simulation
“AlphaEvolve transformed the speed and efficiency of our computational lithography frameworks and, more impressively, demonstrated the potential of these models to design their future selves, all the way down to the atoms.” — James Proud, CEO, Substrate.

Visit the blog to read more about how Substrate applied AlphaEvolve to its computational lithography framework, achieving a multi-fold increase in runtime speed, enabling them to run significantly larger simulations of advanced semiconductors.

WPP: Cracking the code of campaign success
"WPP faced a ceiling in predicting creative campaign performance, as their manual model optimizations yielded only marginal 1% accuracy gains despite significant time and effort. To overcome this challenge, WPP’s Research team utilized AlphaEvolve to autonomously propose, evaluate, and refine candidate model architectures rather than relying on slow manual experimentation. This agentic framework effectively bypassed their trial-and-error limits, successfully navigating complex, high-dimensional campaign data and class imbalances. As a result, WPP achieved a highly significant 5–10% (across different use cases) increase in both prediction accuracy and downstream recommendation scores, outperforming all previous baseline models including neural and fine-tuned Gemma models." — Anastasios Tsourtis, Lead Data Scientist, WPP.

Visit the blog to read more about how WPP used AlphaEvolve to optimize machine learning models for digital marketing campaigns, delivering a 10% lift in prediction accuracy and up to a 7% boost in downstream recommendation scores.

Hardening our own infrastructure and scientific research

Beyond external deployments, Google has integrated AlphaEvolve as a core engine to scale its own state-of-the-art infrastructure. As detailed by Google DeepMind, AlphaEvolve has successfully optimized the silicon design of next-generation Tensor Processing Units (TPUs) with a highly efficient, counterintuitive circuit layout, refined Google Spanner’s Log-Structured Merge-tree compaction heuristics to reduce write amplification by 20%, and reduced software storage footprints by nearly 9% through new compiler optimization strategies. Additionally, the agent has made critical contributions to scientific research, boosting predictive accuracy across 20 natural disaster risk categories by 5%, and discovering quantum circuits with 10x lower error rates for running complex molecular simulations on Google's Willow quantum processor.

According to Pushmeet Kohli, Chief Scientist, Google Cloud & Vice President, Science at Google DeepMind, “AI is moving beyond acting as a productivity assistant that accelerates how we work to a discovery engine that expands what we can achieve. By autonomously navigating complex computational search spaces, tools like AlphaEvolve are helping researchers and engineers uncover breakthrough algorithms that augment traditional human intuition”.

Start evolving your codebase today

Getting started with AlphaEvolve requires only two core inputs on your end:

Seed program: The initial algorithm written as code. You designate which segments of code are open to optimization and provide them to AlphaEvolve
An evaluator: A deterministic client-side evaluation script that compiles, tests, and scores the mutated candidates, returning one or more scalar metrics for AlphaEvolve to maximize.

Your client-side runner queries the AlphaEvolve API to acquire mutated candidate solutions, runs them through your client-side evaluator (which can be running anywhere), and submits the scores back to AlphaEvolve which you sample from.

To use AlphaEvolve we recommend getting going through the documentation. After quickly setting up the AlphaEvolve API using the onboarding guide, we recommend starting going through the repository with the basic colab examples to understand how the AlphaEvolve heuristic works. For agentic workflows, you can easily get started using the AlphaEvolve Skill in your IDE of choice, such as Antigravity or Claude Code. For more complex experimentation, our best practices guide and advanced examples provide additional resources to run through detailed AlphaEvolve experiment workflows.

A developer's guide to publishing agents in Gemini Enterprise and Google Cloud Marketplace

Tue, 07 Jul 2026 16:00:00 +0000

Software-as-a-service (SaaS) is evolving into Agents-as-a-service (AaaS).

Instead of isolated applications, developers are creating AI agents that interoperate using standardized open protocols such as the Agent2Agent (A2A) protocol and can be orchestrated through centralized agent platforms like Gemini Enterprise Agent Platform.

When building for your specific use case, we believe the goal should always be to engineer high-quality agents that combine autonomy with the ability to reliably execute complex, multi-step workflows that deliver clear business value. For agent builders and developers looking to publish and commercialize these high-impact, third-party agents through Google Cloud Marketplace and to deploy them to the Gemini Enterprise app, this guide provides a step-by-step path to a fully integrated, marketplace-ready solution.

Step 1: Design your agent architecture for integration with Marketplace

The end-state architecture bridges Google Cloud Marketplace billing, identity provider (IdP) security, and Gemini Enterprise Agent Platform.

Here’s an overview of these architectural elements:

Customer project: Where users discover agents via the dedicated Agent Marketplace category within Google Cloud Marketplace and interact with these agents through the Gemini Enterprise app.
Partner project: Hosts your agent as well as the marketplace handler, which handles the logic for procurement, and Dynamic Client Registration (DCR) for authorization.
Partner Marketplace project: Manages the Partner Procurement API and Pub/Sub topics for Marketplace events like account creation or entitlement approvals.

Step 2: Review the organizational requirements to sell on Marketplace

Join the Google Cloud Partner Network: If you're new to offering your solutions on Marketplace, join the Google Cloud Partner Network.
Review Agent-as-a-Service listing requirements. Verify that your organization meets the requirements to list your solutions on Marketplace.
Marketplace Vendor Agreement: Review and accept the Marketplace Vendor Agreement (MVA).
Nominate your agent for Google Cloud Marketplace by contacting your Google Cloud representative.

All agents listed on Marketplace must comply with the above standard requirements plus several agent-specific mandates:

Define your agent use case: We recommend defining specific, agentic use cases targeting high-value enterprise functions designed to solve tangible pain points and scale across multiple enterprise customers.
A2A protocol adherence: Agents must comply with the A2A protocol specifications for interoperability. This can include the A2UI protocol which enables your agents to generate rich, interactive user interfaces.
A2A Agent Card: Create an Agent Card, a JSON file declaring capabilities (skills), authentication methods, and service endpoints.
Authentication: Agents must support public access or OAuth 2.0 Authorization Code Grant Flow.
Marketplace integration: Mandatory integration with Procurement APIs and Pub/Sub for entitlement lifecycle management.

Step 3: Review the technical requirements for your agent to be compatible with Marketplace and the Gemini Enterprise app

A2A protocol

When designing and implementing your agent, ensure you follow the A2A protocol documentation. This will guide you on choices for interaction patterns (e.g., streaming or asynchronous tasks) that your agent can provide and can include incorporating an interactive UI experience using the A2UI protocol. Using A2UI allows you to leverage the latest and greatest UX controls available—such as advanced, dynamic charts and modern interaction models. By utilizing these native user controls, you ensure your agent doesn't just function reliably, but looks, feels, and operates with a premium sense of "pride in craft" inside the Gemini Enterprise app.

A2A agent card

To list your Agent-as-a-Service product on the Marketplace, you must provide an A2A Agent Card for your agent. The Agent Card is a JSON file declaring the agent's capabilities (skills), supported authentication & authorization methods, and service endpoints.

The Gemini Enterprise app relies on your Agent Card to:

Display your agent name, description, and other necessary metadata.
Locate endpoints for Dynamic Client Registration (if supported).
Discover agent entry points for sending messages or getting task execution status updates.
Determine the required authentication/authorization methods.

Here is an example Agent Card with definition below.

code_block: <ListValue: [StructValue([('code', '{\r\n "name": "AI Agent Example",\r\n "protocolVersion": "1.0",\r\n "description": "Marketplace agent example.",\r\n "url": $AGENT_APP_URL,\r\n "preferredTransport": "JSONRPC",\r\n "provider": {\r\n "organization": $AGENT_PROVIDER_ORGANIZATION,\r\n "url": $AGENT_PROVIDER_URL\r\n },\r\n "version": "1.0.0",\r\n "capabilities": {\r\n "streaming": false,\r\n "pushNotifications": false,\r\n "extensions": [\r\n {\r\n "uri": "https://cloud.google.com/marketplace/docs/partners/ai-agents/setup-dcr",\r\n "params": {\r\n "target_url": $AGENT_DCR_URL\r\n }\r\n }\r\n ]\r\n },\r\n "defaultInputModes": [\r\n "application/json"\r\n ],\r\n "defaultOutputModes": [\r\n "application/json"\r\n ],\r\n "skills": [\r\n {\r\n "id": "current_time_generation",\r\n "name": "Current time generation",\r\n "description": "Generates a current time.",\r\n "tags": [\r\n "time"\r\n ],\r\n "examples": [\r\n "What time is it?"\r\n ]\r\n }\r\n ],\r\n "supportsAuthenticatedExtendedCard": false,\r\n "iconUrl": $AGENT_ICON_URL,\r\n "security": [\r\n {\r\n "oauth2": [\r\n $AUTH_SCOPE\r\n ]\r\n }\r\n ],\r\n "securitySchemes": {\r\n "oauth2": {\r\n "type": "oauth2",\r\n "flows": {\r\n "authorizationCode": {\r\n "authorizationUrl": $AUTHZ_URL,\r\n "tokenUrl": $TOKEN_URL,\r\n "refreshUrl": $REFRESH_URL,\r\n "scopes": {\r\n $AUTH_SCOPE: $AUTH_SCOPE_DESCRIPTION \r\n }\r\n }\r\n }\r\n }\r\n }\r\n}'), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7f81543fde20>)])]>

$AGENT_APP_URL - A required field representing the base URL endpoint where the A2A agent can be reached. All API calls to the agent will use this as the base path.
$AGENT_PROVIDER_ORGANIZATION - A required field representing the agent provider's organization.
$AGENT_PROVIDER_URL - A required field representing the agent provider's website or relevant documentation.
$AGENT_DCR_URL - A required field if the agent implements Dynamic Client Registration (DCR).
$AGENT_ICON_URL - An optional field providing a URL to an image file to be used as an icon for the agent. If provided, it will be displayed in the Gemini Enterprise app.
$AUTH_SCOPE - An array of strings listing the scope names required for the client to access the agent's operations.
$AUTH_SCOPE_DESCRIPTION - Scope description. Example: "Permission to retrieve email address of the user.”
$AUTHZ_URL - A required part of the OAuth2 security scheme definition for the Authorization Code flow. It specifies the URL of the authorization server's endpoint used to obtain an authorization code from the resource owner. This follows the OpenAPI Specification.
$TOKEN_URL, $REFRESH_URL - URLs for the client to exchange the authorization code for an access token and a refresh token (can be the same).

Authentication and authorization

Implement authentication and authorization for your agent according to the A2A protocol. To allow the Gemini Enterprise app to call your agent, you must establish one of these two methods for your agents:

Public Access: No authentication required. Suitable only for agents that do not access any user data or sensitive resources.
OAuth 2.0 Authorization Code Grant Flow: This is the standard flow for delegated user authorization. Users will be prompted to authorize your agent to access their data or act on their behalf.

Dynamic Client Registration (DCR)

Traditionally, connecting a third-party app to an enterprise system required manual copying of Client IDs and secrets. DCR eliminates this by allowing Gemini Enterprise to programmatically register itself as an OAuth client with your agent's authorization server.

How the DCR Flow Works:

Discovery: The Gemini Enterprise app reads your Agent Card to find the DCR endpoint.
Request: Google sends an HTTP POST to your endpoint containing a software_statement which is a cryptographically signed JSON Web Token (JWT).
Validation: Your backend verifies the JWT signature using Google's public keys to ensure the request is authentic.
Provisioning: Upon success, your server creates a new OpenID Connect (OIDC) application in your identity provider (e.g., Okta) and returns the client_id and client_secret to Gemini Enterprise.

code_block: <ListValue: [StructValue([('code', 'DCR Request\r\n{\r\n "software_statement": "eyJhbGciOiJSUzI1NiIsImtpZCI6ImY1OTIwZDJmMjIyYjNjMTE3Y2MyZmQzZmQxYWJjNzM..."\r\n}\r\n\r\nJWT Decoded\r\nHere is the decoded value of software_statement parameter:\r\n\r\nHeader:\r\n{\r\n "alg": "RS256",\r\n "kid": "f5920d2f222b3c117cc2fd3fd1abc7367fd00402",\r\n "typ": "JWT"\r\n}\r\nPayload:\r\n{\r\n "aud": "https://your-provider.com",\r\n "auth_app_redirect_uris": [\r\n "https://vertexaisearch.cloud.google.com/oauth-redirect"\r\n ],\r\n "exp": 1766773074,\r\n "google": {\r\n "order": "xxxxxxxx-c3bc3976a8e0"\r\n },\r\n "iat": 1766772774,\r\n "iss": "https://www.googleapis.com/service_accounts/v1/metadata/x509/cloud-agentspace@system.gserviceaccount.com",\r\n "sub": "xxxxxxxx-xxxx-xxxx-xxxx-4656e5b81fe8"\r\n}\r\nDCR Response\r\n{\r\n "client_id": $CLIENT_ID,\r\n "client_secret": $CLIENT_SECRET,\r\n "client_secret_expires_at": 0\r\n}'), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7f81543fd040>)])]>

Note: Validating the JWT ensures the request is from Google, but you must cross-reference the google.order ID against your database to ensure the user has actually paid.

Step 4: Publish your agent listing on Marketplace

Once you’ve built your agents, you will need to publish and offer them on Google Cloud Marketplace. This is where you describe your agent and define availability and pricing models. The seller journey begins in the Producer Portal accessible through Google Cloud Console:

Select Solution Type: Choose "AI Agent as a Service" as the product type in the Producer portal.
Upload Agent Card: Provide the Agent Card JSON file via a Google Cloud Storage (GCS) bucket.
Availability: Decide whether the AI agent listing can be purchased through publicly available pricing (self-service) or available via private offer only.
Pricing: Create your pricing plan and choose the pricing model you want to use to monetize the agent through Marketplace.
Technical Integration: Configure the backend procurement. No frontend integration is required for this solution type.
Validation and End-to-End testing: Google Cloud reviews the agent's functionality, security, and pricing model before it is published to the catalog.
Publish: Agent is now successfully published and available in Google Cloud Marketplace.

Step 5: Managing transactions and registrations in Marketplace and the Gemini Enterprise App

There are distinct phases to the procurement and registration lifecycle of agents on Google Cloud Marketplace and the Gemini Enterprise app, which is critical for establishing strict enterprise governance, preventing shadow IT, and ensuring seamless compliance across the organization. A secured chain of custody is managed across three key personas: the Billing Administrator, who maintains financial oversight by controlling procurement and spending on Google Cloud Marketplace; the Discovery Engine Administrator, who acts as the technical gatekeeper by securely registering verified agents and determining organizational access in Gemini Enterprise; and the Discovery Engine User, who can safely leverage the agent's full capabilities within their Gemini Enterprise app only after completing proper identity authorization.

1. Procurement Flow - Async (Google Cloud Marketplace)

Once listed, the backend procurement sequence follows these steps:

Trigger: A customer with Billing Administrator privileges clicks "Subscribe" (for self-serve listings) or accepts a "Private Offer" (for tailored private offer only listings).
Notification: Google sends a Pub/Sub notification to your environment.
Approval and storage: Your integrated marketplace handler approves the account and the entitlement via the Partner Procurement API.
Activation: The handler records the transaction by storing the unique Order ID in a database like Firestore, instantly activating the subscription or offer for the customer.

As shown above, the Billing Administrator executes a one-click subscription to activate the Lovable Agent free plan alongside their already active SaaS subscription procured through Cloud Marketplace.

2. Registration flow - sync (Gemini Enterprise)

After successful procurement, the customer's administrator links the purchase to their actual Gemini Enterprise app environment:

Redirect to Gemini Enterprise: The Discovery Engine Administrator will see a "Go to Gemini Enterprise" option directly on the procured Marketplace listing.
Project Verification: Clicking this prompts the administrator to log into the Google Cloud project where their Gemini Enterprise licenses are allocated. Note that the customer must ensure this destination Google Cloud project is actively linked to the specific billing account used during procurement.
DCR Handshake: The Discovery Engine Administrator configures the agent within the Gemini Enterprise app. At this point, your Dynamic Client Registration (DCR) logic validates the incoming JWT's Order ID against your Firestore records. If the IDs match, the secure registration completes successfully.
Agent successfully Registered: Agent is now successfully registered in Gemini Enterprise. Discovery Engine Administrator can now decide whom to give access to the agent.

Following procurement, the Discovery Engine Administrator registers the Lovable Agent into the Gemini Enterprise app to make it available to authorized users across an organization.

3. End-User Activation Flow (Gemini Enterprise)

Once the agent is securely registered, it becomes discoverable to your target enterprise users:

Gemini Enterprise in-app agent discovery and requests: End users have the ability to browse and directly request access to any available partner-built agent from Cloud Marketplace within the Agent Gallery in the Gemini Enterprise app. When a request is submitted, the Discovery Engine Administrator can review the request and coordinate directly with the organization’s Billing Administrator to procure the agent through Google Cloud Marketplace, and, if already procured and registered, can give access to the end user.
Access: Once access is given to the agent, any end user with an active Gemini Enterprise app account and Discovery Engine User role and license will be able to invoke the agent within their Gemini Enterprise app.
Authorization: Upon the first interaction, the user will be prompted to complete an OAuth authorization by inputting their partner-system username and password. Once authenticated, they can seamlessly leverage the agent's full capabilities from the Gemini Enterprise app chat interface.

An end user seamlessly invokes the Lovable Agent inside the Gemini Enterprise app, completes the one-time partner authorization prompt, and initiates a live conversational task workflow.

An end user requests access to Atlassian Rovo, another agent available from Marketplace, directly from the Agent Gallery in the Gemini Enterprise app. In this demo scenario, the agent has already been procured from Marketplace, allowing the Discovery Engine Administrator to verify, integrate, and instantly grant access.

Get started

Building agents for Gemini Enterprise and Google Cloud Marketplace as an AI Agent-as-a-Service solution provides a path to extend your reach and to get your agent into the daily workflow of millions of enterprise users.

We encourage you to start building today using tools like the Agent Development Kit (ADK) and to learn more about how you can accelerate your growth in the era of the agentic enterprise with Google Cloud Marketplace.

For any assistance, you can contact Google Cloud Marketplace support team.

Report: 83% of organizations need to upgrade their infrastructure to support agentic AI

Tue, 07 Jul 2026 16:00:00 +0000

For years, enterprise AI has been synonymous with conversational AI — the customer service bots and digital assistants we interact with every day. But today, the market has shifted. We’ve officially moved from moving from AI that answers through simple chats, to AI that takes action, automated workflows, and executes complex tasks on its own. While this unlocks entirely new use cases, there’s a catch: it places significant stress on the underlying infrastructure we’ve relied on in the past.

We recently surveyed more than 1,400 senior IT leaders for our State of AI Infrastructure report, and a resounding pattern emerged: the gap between AI ambition and infrastructure reality is widening. In fact, 83% of organizations say they require infrastructure upgrades to support production-grade agentic AI.

Why? Because yesterday’s infrastructure simply wasn't built for agents that act autonomously.

In this blog, we lay out the core insights from our research on how leading organizations are rethinking their infrastructure to build resilient, fluid foundations. For more details and depth, we encourage you to download and read the full report.

Escape the “inference tax” with fluid compute

Agentic workloads introduce a new level of scale, where a single prompt can trigger hundreds of downstream actions, requiring massive context windows to be held in memory. Trying to run these continuous reasoning loops on legacy architecture is financially unsustainable. In fact, 62% of leaders are seeing a significant inference tax driven by data egress fees, storage bloat, and idle specialized hardware. Furthermore, 81% cite operational complexity as a hidden cost of scaling AI.

To fix this, organizations need fluid compute — the ability to dynamically match the right silicon to the right task while minimizing operational overheads.

For heavy training: Compute accelerators like our new TPU 8t deliver tremendous scale to train the world's most sophisticated models.
For low-latency inference: The TPU 8i, meanwhile, was purpose-built to maximize on-chip memory, so agents can think and react in real-time.
For orchestration: General-purpose compute powered by CPUs is emerging as a critical component for driving AI control plane operations. Using highly efficient, Arm-based processors like Google Axion, organizations can cost-effectively run reinforcement learning simulations and orchestrate agents.

Managing agent sprawl with centralized governance

Agents are designed to act autonomously — reading emails, querying databases, and executing workflows across your business. But as agentic AI scales, organizations are facing a new challenge: agent sprawl. How do you manage thousands of autonomous agents scattered across diverse platforms, without losing visibility and control?

It’s no surprise that 79% of tech leaders cite security, governance, and MLOps as their top challenge to scaling inference. In the agentic era, you need a mature governance strategy before you can innovate. This entails creating a centralized control plane that provides a single system of record for agent permissions, identity, and workflows.

Instead of patching together disparate tools, leading enterprises are relying on solutions like Agent Gateway to enforce enterprise-grade governance. Agent Gateway gives you the visibility you need to see exactly how agents are sharing data. It lets you define precise read/write scopes and maintain full audit trails of every interaction, and it provides human-in-the-loop oversight for when an agent needs approval before taking a critical action. This drive for unified, straightforward governance explains why 78% of organizations now source their gen AI solutions directly from their primary cloud partner — a 30 point increase from 2025.

A unified data layer

Agents perform reasoning, meaning they constantly run heavy queries across your organization. If your data is fragmented across silos, your AI is effectively flying blind.

To move from managing disconnected data to gathering unique and actionable business context, leaders are adopting a unified data layer. Using tools like Smart Storage — which automatically annotates unstructured data to make it searchable — and the Cross-Cloud Lakehouse, agents can natively read and understand data no matter where it lives, without needing custom pipelines or duplicated data.

Hybrid multicloud and digital sovereignty

The debate between public cloud and local computing is settled: hybrid is the destination. In fact, 52% of organizations now use a hybrid multicloud architecture.

For technology leaders, this shift is largely driven by digital sovereignty and data gravity. Indeed, 48% of leaders are prioritizing infrastructure with strict data residency controls. You need the flexibility to run AI where it complies with shifting local laws. Whether that’s leveraging the public cloud for broad compute, or bringing foundational models entirely on-premises via Google Distributed Cloud for air-gapped isolation, modern infrastructure must adapt to geopolitical realities, not the other way around.

AI at the edge

For technology leaders and infrastructure architects, relying on a strictly centralized cloud topology to process every agentic interaction is not a viable strategy. A staggering 90% of organizations now rank edge deployment as important for AI initiatives, with 72% describing it as extremely or very important.

Moving AI to the edge solves three issues:

The latency bottleneck: Real-time agents — especially those that rely on voice, video, or financial trading algorithms — can't afford the microsecond gap of a round-trip to a distant data center.
Operational resilience: If an internet connection drops, business can't stop. Edge deployment ensures that agents running in manufacturing plants, retail stores, or hospitals can continue functioning autonomously.
Sustaining cost-efficiency: Running always-on, continuous reasoning in the cloud is expensive. By utilizing highly optimized models on edge devices (like smartphones, IoT devices, or local warehouse servers), organizations shift the compute burden locally, drastically cutting variable per-token costs.

Breaking through the energy wall

Energy consumption used to be a sustainability metric reserved for annual reports. Today, it plays a crucial operational role. 91% of leaders now factor power consumption into their hardware selection, with 61% rating it as a primary or significant factor.

For technology leaders, power consumption presents a three-fold barrier to growth:

Grid scarcity: You simply cannot buy more power in certain regions, heavily limiting how much compute infrastructure can be provisioned.
Regulatory compliance: Energy efficiency is now a strict legal prerequisite to operate. For example, in Germany, new data centers must achieve a Power Usage Effectiveness (PUE) of 1.2 or lower. And Ireland now mandates that large data centers provide 100% on-site dispatchable generation to match their grid draw.
Infrastructure economics: Inefficient power envelopes drastically inflate the Total Cost of Ownership (TCO) of AI deployments. Accommodating high-power hardware requires massive capital expenditure (CapEx) for advanced cooling architectures, specialized rack designs, and facility upgrades.

To address the energy wall, technology leaders must treat energy as a strategic asset. One of the focus areas for optimization must shift to performance-per-watt. This is why co-designed silicon is becoming so important. For example, our new TPU 8t delivers nearly three times the performance of the prior generation while being up to twice as energy-efficient.

Unified, AI-optimized infrastructure

Ultimately, you cannot solve the challenges of tomorrow’s agentic systems with yesterday’s architecture. When engineering teams are forced to manually integrate heterogeneous compute, storage, and networking layers, organizations incur high operational overhead just to ensure basic interoperability.

To innovate quickly and cost-effectively, technology leaders are therefore moving toward holistic, unified systems.

This is the philosophy behind Google Cloud’s AI Hypercomputer. It’s an architecture where every layer is co-designed and co-engineered to work together. The custom silicon (TPUs, GPUs, CPUs) isn't designed in a silo; it's engineered alongside the ultra-high-bandwidth networking (Virgo Network), the storage (Managed Lustre, Hyperdisk), and the software orchestration layer (GKE).

Bridging the digital and physical worlds

When you embrace this co-designed, holistic approach, the results go far. With this level of scalable, fluid intelligence operating at the edge, we're entering the era of physical AI. A new generation of autonomous robots can sense, simulate, and navigate the physical world, practicing tasks millions of times in digital twin simulations on Google Cloud before they ever set foot in the real world. From performing complex industrial inspections to capturing cinematic videography, AI is now solving tangible problems in the real world.

Your blueprint for agentic AI

Adapting your infrastructure to meet the demands that agentic applications place on your systems will help you move from pilot to production. The organizations set to thrive in 2026 are embracing a unified foundation that is cost-efficient, resilient at the edge, optimized for autonomous action — and governed by default.

Ready to start? Download the State of AI Infrastructure report to explore the data behind our findings, and discover how your peers are already building for success.

20 questions for the Agentic Enterprise (and how Agent Platform can help)

Tue, 07 Jul 2026 16:00:00 +0000

If you’re an IT leader, you might be getting a lot of questions about how to build and deploy agents.

The pressure to move fast is intense, but the engineering reality is incredibly complex. Where do your teams even begin? How do you untangle a fragmented mess of disconnected tools? And as things grow, how do you ensure your agents don’t accidentally leak sensitive data, or burn through your token budget in an afternoon? It’s a lot to balance, and trying to establish a secure foundation for an entire organization can quickly feel overwhelming.

That’s why we built Gemini Enterprise Agent Platform. It gives your technical teams a unified destination to build, scale, govern, and optimize both customer-facing agents and the ones managing your internal operations. Agent Platform handles the underlying complexity so your teams can focus on driving actual business value.

To help you navigate these conversations, we gathered 20 essential questions to ask your engineering teams, along with some practical advice and code examples to get you going. Let’s dive in.

The build phase — establishing the foundation

#0 Who is building the application?
Before choosing a tool, look at who on your team is actually doing the work. Is it your engineer? Your legal team? Building with AI is no longer exclusive to high-code engineers. Anyone can vibe code now. This incredible accessibility has turned millions of non-coders into creators who can build and launch applications in seconds. So it means your work could be coming from anywhere. This may sound like an obvious step, but it’s an important one in the AI era.

The ecosystem now spans a spectrum of personas: no-code business experts defining logic via visual interfaces (think: your business teams, sales, and marketing), low-code developers assembling modular parts, and high-code engineers creating bespoke, custom reasoning loops. Successful adoption means choosing a platform that empowers all three personas without siloing your data or security.

#1 Where should my developers start?
When setting up an agentic strategy, it's easy to focus exclusively on the end product, like the agents that will handle customer support or financial analysis. But to build those sophisticated agents, you have to start by empowering the builders who write their underlying logic. Your developers need their own specialized AI tools, like coding agents, to accelerate code generation, scaffolding, and integration.

However, most coding agents are isolated. They can only analyze the immediate file they are working on, with no connection to your live databases, internal documentation, tech stack, or business systems.

To keep your devs moving quickly without sacrificing governance, we recommend using Google Antigravity as your primary engineering harness, and then integrating specific extensions based on what that team is building. Here’s a helpful breakdown:

For core application engineers: Use the upgraded Agent Development Kit (ADK) as your baseline framework, paired with Agents CLI to handle the entire agent lifecycle from the terminal.
For data engineers: Plug in the Google Cloud Data Agent Kit, which provides dedicated skills and Model Context Protocol (MCP) tools tailored for data pipelines.
For Google Cloud ecosystems: Deploy Agent Skills to give your coding environment native capabilities across Google products.
For integrated IDE experiences: Connect the Developer Knowledge Base via MCP to stream official documentation directly into your teams' workflows.

#2 Who are we building for? Humans, or other agents?
Like #1, this might sound like a straightforward question, but you'll want to decide early on if you're building an AI agent for your employees to talk to directly, or if it's meant to coordinate with other agents behind the scenes. Your design requirements will look completely different depending on who — or what — is interacting with the system, so keeping everything under your team's control starts with knowing exactly who you're building for.

If your answer is humans (building for employees or customers), focus on user experience. You can host and share these tools in a single place like the Gemini Enterprise app, or use the Agent-to-User Interface (A2UI) framework to drop interactive components directly into your custom apps.

If your answer is agents (and you’re building agents meant to talk to other agents), focus on interoperability. By adopting the open Agent2Agent (A2A) protocol, an open standard for seamless communication and collaboration between AI agents, your agents can use standardized metadata to discover each other, pass context, and securely delegate background work across completely different enterprise frameworks.

See question #14 for more on user and agent identity.

#3 Which agent development tool should I use?
With so many frameworks available, it’s easy for engineering teams to default to fragmented, homegrown setups. To simplify this, we look at agent development as a four-rung ladder, which allows teams to slide between out-of-the-box configuration and code-first control:

Rung 1: Agent Studio (low-code): A visual workspace inside Agent Platform for rapid prototyping and business teams. Build an agent with Agent Studio here.
Rung 2: Managed Agents API (Agent-as-a-Service): For technical teams who want to define agent behavior via API and let Google handle the infrastructure inside a secure sandbox. Build a custom agent with Managed Agents API here.
Rung 3: Antigravity 2.0: A dedicated workspace for developers leveraging AI for advanced coding tasks and engineering pipelines. Build with Antigravity here.
Rung 4: Agent Development Kit (ADK 2.0): An engineering-first, code-first framework for software engineers building highly custom, multi-agent networks from scratch. Build a sample agent with ADK here.

#4 Should I start with one agent or many, and how do I specialize them?
Always advocate for your teams to start with a single, highly specialized agent for initial prototyping. If an agent tries to do everything, a few things might happen: accuracy drops, latency spikes, and debugging becomes a nightmare. To avoid this, write tight instructions and limit the tools it can access.

As your workflows grow more complex – or if you hit model context limits – have your engineers graduate to a multi-agent system. At its core, a multi-agent system is a collection of individual, autonomous agents that collaborate to achieve a goal. Using a framework like ADK, they can organize agents into a network of sub-agents where a coordinator delegates specific tasks to specialized team members, maintaining clear organizational logic.

Example: Build a sample multi-agent solution with ADK here.

The scale phase - connectivity and interoperability

#5 How do we connect enterprise data and maintain the right business context?
Your agents need access to the right data to be truly useful. This is where “enterprise truth” comes in – it's what we call your enterprise’s specific data, tools, constraints, policies, and processes that the agent needs to be successful.

While there are several ways to accomplish this, the emerging practice is using open standards like Model Context Protocol (MCP) to connect your agents directly to live databases and business apps. However, simply establishing connectivity isn’t enough. To help your agents work accurately and avoid hallucinations, you must also organize this data with clear business context, metadata, and logic. This structured approach ensures your agents don’t just pull raw information, but actually interpret it correctly to and make better decisions across your organization

Example: Build a sample multi-agent solution with ADK and MCP Toolbox (Managed Server).

#6 How do we connect agents built on completely different frameworks?
In a large organization, different teams will naturally build agents using the tools that best fit their specific needs, whether that's LangGraph, a homegrown framework, or something else entirely. However, if these systems can’t communicate, you end up with isolated data and workflow silos.

Establishing a universal communication standard allows agents developed on completely different platforms or frameworks to exchange intents, state, and results without specialized integration work.

For cross-framework connectivity (e.g., connecting a LangGraph-based HR agent to an ADK-based CRM agent ), you can implement the A2A protocol. This allows a microservices-style communication pattern across multiple distinct agents, ensuring they can securely talk to each other.

Example: Build a sample multi-agent solution with ADK and A2A.

#7 How do we help agents find the specific tools they need?
Stuffing an agent's context window with multiple tools and APIs degrades performance, increases latency, and drives up token costs. Just as we use RAG to dynamically fetch data on demand, we must apply the same dynamic retrieval strategy to agent tooling.

By utilizing focused agentic Skills, agents load capabilities only when a task requires them. Instead of parsing a massive library of generic instructions, the agent pulls from a single, task-specific index card—ensuring precise, tightly controlled execution.

Example: Build a multi-agent orchestration pattern with ADK and Skills

#8 How do we deploy our agents so they can easily scale?
This is the million-dollar question — or, perhaps more accurately, the "tokens-per-minute" question. The key isn't just about choosing the cheapest option, but about finding the right recipe of tools and services that aligns with your workload patterns.

To scale your agents without racking up massive infrastructure overhead, your teams should deploy agents within a fully managed, serverless execution environment. Agent Runtime is a set of services that enables developers to deploy, manage, and scale AI agents in production. Agent Runtime handles the infrastructure to scale agents in production so you can focus on creating applications.

A production-ready runtime must offer elastic auto-scaling to handle sudden usage spikes, containerized flexibility to bundle custom software dependencies, and native support for bidirectional streaming to ensure low-latency, real-time interactions. The architecture must also integrate built-in private networking interfaces to securely connect to internal enterprise data without public internet exposure.

Example: Build a multi-agent orchestration pattern with Agent Runtime

#9 What if our agents lose track of context during long-running tasks?
To help your AI agents work more effectively, you can give them both short-term and long-term memory. This means using real-time session state to keep immediate conversations going, and a long-term storage layer to remember user preferences and past interactions — all while keeping your agents safely under your team's control.

Agent Platform handles this across two layers. In ADK, a sessionService handles the immediate steps of a multi-stage task, while Agent Memory Bank acts as a persistent, long-term storage layer to recall past user preferences and project outcomes over time.

Example: Build a multi-agent orchestration pattern with Agent Memory Bank and ADK memory

The optimize phase — trust and efficiency

#10 How do we limit the blast radius for an agent running scripts or using a browser?
If your agents need to run Python, execute scripts, or browse the web to gather data, they shouldn’t do it directly on your network. Running these tasks in a temporary, isolated sandbox environment makes it easy to isolate any untrusted code or runtime logic errors, keeping them completely separate from your core enterprise systems.

Plus, using our agent runtime with a built-in sandbox helps protect your primary infrastructure and lets your agents safely execute tool calls under your team's full control.

Example: Build a multi-agent orchestration pattern with Agent Sandbox

#11 How do I ensure my agent stays on-brand?
An agent represents your corporate identity. While defining system prompts with clear constraints is a starting point, relying on prompts alone is insufficient (and risky) for production security. You need a mandatory safety layer that enforces core corporate rules and tone constraints, making sure the agent remains bounded regardless of the model's inherent probabilistic nature.

Guardrails turn an unpredictable LLM into a safe enterprise system by allowing the agent to have the flexibility to make autonomous decisions, while keeping it incapable of violating core safety rules. Because these boundaries are implemented as deterministic constraints outside of the agent’s reasoning, they cannot be bypassed.

Complementing this safety layer, structured workflows drive even greater predictability by breaking complex tasks into deterministic, step-by-step pipelines that use code-level routing, conditional logic, and state management to guide the agent through repeatable paths.

Example: Build a multi-agent orchestration pattern with Guardrails agent. See also ADK Workflows

#12 How do we trust the result?
Trust is earned through evidence gathered from rigorous testing and ongoing evaluations across the agent’s entire lifecycle. It’s not an automatic given, but rather a result of your method.

At scale, evaluation is automated using a mix of metrics, human-in-the-loop oversight, and LLM-as-a-judge patterns. By using a more capable model or a specialized self-evaluation agent to audit the primary agent’s output before it reaches the end-user, you can systematically catch inaccuracies and protect the user experience.

Example: Build a multi-agent orchestration pattern with LLM-as-a Judge and Self Evaluation agent built in

#13 How do I control costs that are going overboard?
High-performance reasoning is powerful, but it isn’t cheap. To optimize your spend, try using a tiered approach: employ fast, lightweight models (like Gemini Flash) for high-speed, low-complexity tasks, leverage open source models (like Gemma), and reserve your largest, most expensive reasoning models for final decision-making.

For high-volume production, switch to Provisioned Throughput (PT). Think of it like booking dedicated capacity for your steady, predictable everyday traffic, while unexpected spikes safely overflow into standard pay-as-you-go billing.

You can further protect your budget by trimming context windows with precision RAG, utilizing context caching, setting hard stops on agent iterations, and transitioning predictable parts of the agent workflow into deterministic code where feasible.

Example: Build with cost control in mind and KPIs that matter

The govern phase — security and oversight

#14 How do I align an agent’s data access to match that of its human user?
Aligning an agent’s data access starts with establishing a secure agent identity, which supports three models: operating directly under a user's identity, using the agent's own independent identity, or acting via delegated authority.

For many employee-facing workflows, leveraging delegated authority is the most secure approach. The agent automatically inherits and respects the existing permissions of the employee interacting with it. This guarantees that the agent cannot access data it isn't explicitly authorized to see, eliminating the need to rebuild complex permission structures from scratch while maintaining a clean audit trail.

Example: Build with Agent Identity

#15 How do I manage shadow AI and agent sprawl?
Unmonitored agents create severe data fragmentation and compliance risks. To prevent sprawl, you can use a central agent registry — a single, discoverable directory that automatically inventories every active agent, its business owner, its target dataset, and its permitted tools. Moving away from manual tracking spreadsheets gives your teams visibility into internal AI projects, ensuring that redundant agents are consolidated and orphaned endpoints are safely decommissioned.

Example: Build with Agent Registry

#16 How do I define how users, agents, data, and tools are allowed to interact?
Scaling enterprise automation safely requires a dual-layered policy architecture. First, apply IAM policies to set clear boundaries so agents only access authorized tools and specific data buckets. Second, implement semantic policies that analyze the natural language intent of a user prompt in real time, validating that the agent's planned response aligns with core business rules and compliance mandates before execution.

Example: Build with Policies

#17 How do I enforce those policies and gain visibility into agent activity?
Policies are meaningless without runtime enforcement and a clear audit trail. To achieve this, you need to route all agent traffic through an agent gateway — the network entry and exit point for all agentic interactions. This gateway should automatically intercept calls between users, agents, and tools to instantly block policy violations, sanitize content, and prevent prompt injections. For total visibility, this gateway must generate network-layer telemetry for every single interaction, feeding real-time behavioral metrics and execution traces directly into your observability dashboards.

Example: Build with Agent Gateway

#18 How do I protect prompts and responses against data leakage, prompt injections, and offensive content?
When integrated with Agent Platform, Model Armor intercepts prompts before they reach Gemini models, and intercepts responses before your application receives them.

Based on your configuration, Agent Platform calls the Model Armor service, which inspects or blocks traffic that violates your defined policies — enforcing security measures like prompt injection and jailbreak detection, responsible AI filters, and sensitive data protection. You can configure this integration either by using floor settings for project-level protection or by using templates for per-request protection.

Example: Build with Model Armor

#19 How do I know if something has gone wrong with one of my agents?
To protect your systems, you need to look for behavioral anomalies by auditing your agent's decision-making loop in real time. Running this continuous behavioral audit alongside threat detection ensures you instantly catch whenever a compromised agent attempts a high-risk, uncharacteristic action.

This is where Agent Platform Threat Detection (part of Security Command Center) comes in. If an agent attempts unauthorized database commands or connects to unverified external network addresses, the system flags the event in near-real time for rapid isolation — keeping your automated workforce safely under control.

Example: Build with Threat Detection and Agent Anomaly Detection

#20 How can I manage the complete agent lifecycle in one place?
Nobody wants to click through five different cloud consoles just to push an update or run a test. With Agent Platform, you can simply give your coding agents the specific skills and commands needed to build, scale, govern, and optimize production-ready agents.

We recommend using the Agents CLI in Agent Platform as the central command tool for your development teams. It acts as a direct bridge between local terminal work and live production management, making it much easier for developers to transition from testing to a live launch.

It also allows your teams to version-control agent configurations, run automated evaluations, and seamlessly push updates through your existing CI/CD pipelines. Because the underlying tools and skills are built and rigorously tested by Google's experts, your team can deploy with confidence without having to reinvent the wheel or break their day-to-day coding workflows.

Example: Build with Agents CLI here

Get started today

By tackling these 20 questions early, you can build agents that actually do real work for your business — without keeping your security and operations teams up at night.

Get started with Gemini Enterprise Agent Platform here.

Drive proactive security, prioritize risks with Google Threat Intelligence and Wiz ASM

Tue, 07 Jul 2026 16:00:00 +0000

Being more proactive continues to be a leading goal for security organizations. As AI accelerates the pace of vulnerability discovery and exploitation, organizations will rely on the personalization of their security investments to help prioritize their defenses.

To help you be more proactive by matching your real-world exposures with real-time adversary activity, we’ve begun integration efforts between Google Threat Intelligence and Wiz Attack Surface Management (ASM).

By connecting exposure and validated exploitable risks directly to real-time threat intelligence, we can help you detect and prioritize external-facing exploitable issues and uncover logic-driven vulnerabilities with AI scanning at the speed needed for today’s defenses. This allows you to shift to a strategy that prioritizes actions based on the real-world threats that pose the greatest risks to your organization.

Combining these two perspectives on threats can help you move from reactive maintenance to a proactive security strategy. In addition to detecting your exploitable exposures, you gain insight into which of those exposures are being actively targeted by adversaries.

We will continue to build towards native integration that will feed exposure data directly into the Google Threat Intelligence correlation engine. This automated connection will help you focus on the exposures adversaries are targeting in the wild, and use our real-time threat intelligence to prioritize remediation efforts and threat hunting activities.

Building a proactive security strategy

Google Threat Intelligence provides global visibility into how adversaries operate, tracking their infrastructure and campaign activity in real time.

Wiz ASM maps your external attack surface across cloud, AI, software-as-a-service (SaaS), and on-premises environments to reveal exposed assets like domains, IPs, and APIs. It scans for exploitable vulnerabilities, misconfigurations, and default credentials to validate exploitability. It also scans for and validates exposed secrets and sensitive data.

At the same time, the Wiz Red Agent scans exposures with AI to uncover complex, logic-driven vulnerabilities by reasoning about applications behavior.

The power of this combination lies in the ability to prioritize and hunt with confidence:

Prioritize based on real-world activity: With the incoming integration, exposure data feeds into the Google Threat Intelligence engine. This helps you spot the exposures that adversaries are currently exploiting, allowing your team to focus remediation efforts where they are needed most.
Understand attacker behavior: When a critical risk is flagged, we plan to provide behavior-based guidance alongside the alert. This details how an attacker typically acts after exploiting a vulnerability, using specific host commands or malware, giving your defenders the context they need to hunt for active footprints inside your network.
Discover complex vulnerabilities: The Wiz Red Agent uses AI to scan for logic-driven vulnerabilities, such as authentication bypasses, business logic flaws, and multi-step attack chains, helping you uncover risks that traditional scanners often miss.

Get started with proactive defense

This combined approach is designed to help you streamline your security posture by reducing the noise and focusing on the signals that represent real danger to your organization. To get started with Google Threat Intelligence and Wiz ASM today, contact your Google sales representative.

Shift into high gear with agents: Securing the software-defined vehicle

Mon, 06 Jul 2026 16:00:00 +0000

The automotive industry is at a pivotal crossroads as it hits the gas on adopting new technology. The era of the traditional connected vehicle has shifted into the age of the software-defined vehicle (SDV), notable for rapid innovation with many new capabilities delivered over the air.

By integrating AI and agents, the next generation of SDVs will be capable of turning raw telemetry into actionable insights in real-time, allowing for a fundamental rethink of how vehicles interact with their environment and their users. To better support and secure SDVs, Google Cloud and Valtech have partnered to develop Nexus SDV, a highly-scalable, AI-enabled connected vehicle platform built on Google Cloud. This modular, developer-friendly and open-source solution is designed to manage up to 100 million devices, and features deep integration with Android Automotive OS (AAOS) to streamline data flows and in-vehicle experiences.

We are proud to announce the first release of the Nexus SDV open-source core that showcases how it can reduce total cost of ownership through Arm-based compute and Bigtable, while providing a AI-native environment for building the next era of automotive intelligence.

AI-driven experiences with Nexus SDV

Nexus AI serves as the platform’s intelligent engine, transforming the vehicle from a passive data source into a proactive, agentic partner. Using Gemini models and Gemini Enterprise Agent Platform, Nexus AI can analyze complex telemetry in real-time to provide information for autonomous decision-making and hyper-personalized driver assistance, effectively acting as an intelligent agent that anticipates user needs.

Crucially, this advanced intelligence is paired with a focus on significant total cost of ownership (TCO) reduction. By using high-efficiency Arm-based compute and Bigtable-optimized data storage, the platform lowers the operational costs associated with processing massive data volumes. This modular, AI-native architecture ensures that manufacturers can scale their fleet intelligence rapidly without the prohibitive cloud and development expenses traditionally associated with next-generation vehicle software.

Cloud-native under the hood

The architecture of Nexus SDV is built on a modular, cloud-native foundation designed to bridge the gap between the vehicle edge and the data center. Deep compatibility with AAOS is the keystone of the close integration between the cloud and the vehicle, and will help ensure that high-fidelity telemetry is ingested and synchronized in real-time. This robust data loop allows Nexus AI to quickly push intelligent updates and services back to the vehicle.

By providing this developer-friendly, open framework, Nexus SDV enables manufacturers to manage the entire lifecycle of a SDV with the scalability and reliability of the Google Cloud ecosystem.

Architecture for Nexus SDV.

Defense in depth with Google Cloud Security controls

By building on Google's secure foundations, including secure-by-design and Zero Trust architecture, Nexus SDV supports the heavy lifting of compliance and threat protection. To achieve this, the Nexus SDV architecture implements a comprehensive, defense-in-depth security model across six key elements:

Mutual TLS (mTLS) and public key infrastructure (PKI)
Nexus SDV relies on cryptographic trust chains to authenticate vehicles before any data exchange can occur. The infrastructure uses Google Cloud Certificate Authority Service (CAS) to manage distinct CA pools (server, factory, and registration CAs), ensuring a highly available and secure root of trust.

Specifically, the registration server enforces registration by forcing clients to present a valid "factory-issued" certificate during the initial TLS handshake, extracting and parsing the certificate directly from the connection stream to definitively prove the vehicle's identity. During registration, the server performs Certificate Signing Request (CSR) validation sent by the vehicle before issuing a new operational certificate.

Identity and access management
The system uses identity brokering where Keycloak is deployed as the central OpenID Connect (OIDC) identity provider. Vehicles authenticate against Keycloak using their operational certificate via mTLS to receive a short-lived JSON Web Token (JWT).

For fine-grained access control, a custom NATS Auth Callout service provides dynamic subject permissions: It intercepts all messaging broker connection attempts, validates the Keycloak JWT using public JWK keys, and programmatically maps the vehicle's roles to specific NATS subjects.

For secure service-to-service communication, it uses Workload Identity Federation so pipelines exchange GitHub OIDC tokens for temporary Google Cloud access, removing static credentials, while GKE Workload Identity allows Kubernetes Pods to access backend services like Bigtable by binding Kubernetes service accounts to Google service accounts.

Security is reinforced through restricted IAM scopes, ensuring dedicated service accounts are provisioned with minimal permissions, such as the data API being restricted only to reading from Bigtable. Using VPC-SC, Organization policy constraints and Private Service Connect (PSC) in your deployment context also helps you achieve secure foundations.

Secret management
Nexus SDV relies on centralized secret management to protect sensitive information. All sensitive configurations, database passwords, and cryptographic signing keys are generated dynamically during Terraform infrastructure provisioning and locked inside Google Cloud Secret Manager.

A secret fetching during deployment is used to avoid baking secrets into application code and container images. Instead, services pull signing keys and credentials directly into memory only at runtime, minimizing exposure both at rest and in transit.

Network isolation
To enforce network isolation, the underlying computer infrastructure is heavily shielded. Nexus SDV runs on private GKE clusters where worker nodes have no public IP addresses, preventing direct internet exposure. Additionally, the Keycloak PostgreSQL database uses Cloud SQL IAM Authentication, which allows the Cloud SQL Proxy to connect securely using IAM roles rather than relying on static database passwords or managing IP allowlists.

Secure AI Framework
Google Cloud secures these advanced AI capabilities through a comprehensive, enterprise-grade framework that prioritizes data privacy, model governance, and safe execution, based on guidance from the Secure AI Framework (SAIF). With Gemini Enterprise Agent Platform, security and governance are natively embedded into the machine-learning lifecycle through capabilities, such as dedicated Explainability and Safety controls, continuous Evaluation and Monitoring, and secure model registries.

You can learn more about how we secure AI here.

Data API
Instead of allowing downstream applications and external clients direct access to data stores like Bigtable, Nexus SDV routes data retrieval through a custom Data API. This microservice acts as a secure abstraction layer that translates strictly, such as querying specific vehicle IDs, sensor data types, and predefined time windows, into heavily constrained Bigtable row-range scans and column filters.

By doing so, it serves as a secure gateway that enforces structured data access patterns.

Start your journey with Nexus SDV

Nexus SDV represents a new era of automotive intelligence, delivering an agentic, secure, and cost-efficient platform that empowers manufacturers to harness the full power of AI in an open-source framework. You can learn more about how we are redefining the software-defined vehicle here.

AlloyDB AI Functions - now with revolutionary performance boosts and cost savings

Wed, 01 Jul 2026 18:00:00 +0000

AlloyDB is an AI-native database—it isn’t just a passive data store, it intelligently understands and processes your data. With AlloyDB, you get industry-leading vector and hybrid search, near 100% accurate natural language-to-SQL capabilities to build conversational agents, tools to enable you to build with your agentic IDEs of choice, and the ability to bring the intelligence of foundation models like Gemini directly to your data through AI functions.

In this blog post, we discuss the massive breakthroughs in AI function processing alongside a suite of brand-new AI functions.

But first: what exactly are AI functions? They bring Gemini’s world knowledge to your AlloyDB data. Consider the challenge of managing raw user feedback: it’s unstructured, and difficult to parse through. Before this data can be leveraged for search, it may require pre-processing and entity extraction. Rather than maintaining complex custom pipelines for knowledge extraction, you can use Gemini’s generation capabilities directly within AlloyDB to transform raw text into structured, searchable insights. For example, here is how you can use ai.generate to instantly turn raw feedback into clean, structured JSON (see more examples here):

code_block: <ListValue: [StructValue([('code', "SELECT\r\n log_id,\r\n raw_content,\r\n -- Use Gemini 3.0 to reason through the raw user feedback and extract structure\r\n ai.generate(\r\n model_id => 'gemini-3.1-pro-preview',\r\n prompt =>\r\n 'Analyze this raw customer feedback entry. Extract the country, service name, and a 1-sentence summary of the feedback. Return as JSON.'\r\n || raw_content) AS structured_feedback\r\nFROM raw_feedback_logs\r\nWHERE user_type <> 'internal';"), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7f8142bb21f0>)])]>

Here is a sample result:

log_id	raw_content	structured_analysis
1001	2025-12-16 08:00:01 [ERROR] Service: OrderSvc \| DbConnectionTimeout: Failed to acquire connection from pool "primary-shard-04" after 5000ms.	{"errorCode": "DbConnectionTimeout", "serviceName": "OrderSvc", "rootCause": "The service failed to acquire a database connection from the primary shard pool within the 5000ms timeout limit."}
1002	2025-12-16 08:05:12 [WARN] Service: IdentityProvider \| 401 Unauthorized: Bearer token validation failed for user_id=9942. Signature mismatch.	{ "error_code": "401", "service_name": "IdentityProvider", "root_cause": "The bearer token validation failed due to a signature mismatch." }
1003	2025-12-16 08:12:45 [CRITICAL] Service: AnalyticsEngine \| OutOfMemoryError: Java heap space. Allocation of 1.2GB array failed. Heap usage 99%.	{ "error_code": "OutOfMemoryError", "service_name": "AnalyticsEngine", "root_cause": "The service exhausted available Java heap memory attempting to allocate a 1.2GB array." }
1004	2025-12-16 08:25:33 [ERROR] Service: WebFrontEnd \| 404 NotFound: Resource /api/v3/users/profile/settings not found. Upstream returned 404.	{ "error_code": "404", "service_name": "WebFrontEnd", "root_cause": "The requested API resource for user profile settings was not found by the upstream service." }
1005	2025-12-16 08:35:50 [WARN] Service: NotificationGateway \| GatewayTimeout: External provider "SendGrid" failed to respond within 30s. Retry scheduled.	{"error_code": "GatewayTimeout", "service_name": "NotificationGateway", "root_cause": "The external provider SendGrid failed to respond within the 30-second timeout limit."}

More functions to summarize and analyze sentiment

Our core AI functions —ai.generate, ai.rank, ai.if, and ai.forecast—are now Generally Available. To learn more about use cases for the first three, refer to this blog post. To explore the forecast function in action, check out this deep dive.

Building on this momentum, we have introduced three brand new functions: ai.summarize, ai.agg_summarize, and ai.analyze_sentiment.

ai.analyze_sentiment: Automatically classifies the emotional tone of text as positive, negative, or neutral.
ai.summarize: Condenses lengthy text into its most essential information while preserving the original tone and nuance.
ai.agg_summarize: An aggregate tool that processes multiple rows within a column to generate a single, unified summary for an entire group (e.g., via a GROUP BY clause).

Here’s an example of how to use ai.agg_summarize to consolidate a product reviews for products on a retail website:

code_block: <ListValue: [StructValue([('code', 'SELECT productname, ai.agg_summarize(review) as reviews_summary\r\nGROUP BY productname;'), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7f8142bb2ac0>)])]>

Here is a sample result of summarized reviews for two gaming console products:

productname

reviews_summary

AlphaCore Console

Users praise the stunning 4K graphics, smooth 120Hz frame rates, and the highly ergonomic controller design.

However, several reviews express frustration over the loud cooling fan noise during extended gaming sessions.

Overall, it is considered a top-tier console despite minor thermal and noise complaints.

NeoCore Console

Customers love the exceptional battery life and vibrant OLED display for handheld gaming on the go.

A significant number of users noted that the UI can feel sluggish and the game library is currently limited.

It represents great value for casual gamers but power users may find the performance lacking.

The power of LLMs on your data: now significantly faster and cheaper

We now have achieved unprecedented performance and cost breakthroughs in AI function processing. Previously, running a foundation model call for every single row in a massive database introduced cost and latency constraints. We have shattered these barriers by introducing two breakthrough capabilities:

Smart Batching for AI Functions: This AI Function Acceleration capability provides intelligent batching of AI function calls for optimal performance and quality. This efficiency is achieved by deduplicating prompt overhead; the LLM's boilerplate instructions are transmitted once per batch rather than repeated across every individual row. A question you may have is - “Why not do this in my own application layer?”. That’s because, AlloyDB intelligently determines the right batch size for optimal results - if you underestimate the batch size, you won’t reap gains for cost and latency, and if you overestimate the batch size, the prompt to the LLM could get bloated and lead to hallucinations, or you could exceed the model's token limits. In addition to calculating the perfect batch size for every request, AlloyDB also handles retries automatically out of the box, ensuring your pipeline stays resilient. We did some testing internally and saw massive gains; for example, an up to 2,400x performance boost (processing 10,000 rows/sec) over traditional row-at-a-time LLM calls. This is currently available for the ai.if and ai.rank functions, with support for additional functions coming in the future.

Let’s look at an example of using Smart Batching / Acceleration with ai.if to solve this use case: Imagine a customer on a gadget retail site searching for a camera that can handle an underwater depth of '60 meters or deeper.' Traditional hybrid search will pull the closest semantic and full-text matches, but it misses the hard constraints of numerical data—meaning it might serve up a camera that works only at 20 meters depth. By using AlloyDB’s ai.if-based intelligent filtering, the database actually understands the nuance of depth and makes the query return products that meet or exceed that 60-meter depth criteria. Notice how, in the example below, you don’t need to specify the batch size - AlloyDB handles all the optimizations under the hood when using ai.if.

code_block: <ListValue: [StructValue([('code', "-- Smart Batching / AI Function Acceleration \r\nSET google_ml_integration.enable_ai_function_acceleration = on;\r\nSELECT productid, productname, category,description\r\nFROM products AS p\r\nWHERE\r\n ai.if(\r\n 'Evaluate if the product description indicates that the product is waterproof at depth 60m or deeper. Description:'\r\n || description);"), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7f8142bb2bb0>)])]>

Here is a sample result on a hypothetical gadgets site. Notice how the expanded descriptions of products really match the criteria of working at a depth of 60 meters:

Optimized AI Functions: For even greater efficiency, we’ve introduced an optimized mode, starting with ai.if. By deploying a small, proxy model that utilizes your embeddings and is trained on your specific LLM outputs, we can process decisions natively within the database. This drastically reduces the need to call the external LLM - and based on some of our internal tests, we saw staggering gains; for example, up to 100,000 rows processed per second (a 23,000x improvement) and costs slashed by 6,000x (down to 1/10th of a cent). For technical insights on this technique, including when it works best and when not, refer to this blog post. AlloyDB does the following when using optimized ai.if:

Trains a proxy model: AlloyDB trains a lightweight proxy model on a sample of your data. This happens in the background when you use the PREPARE statement with ai.if function to train the model for optimized queries.
Executes the query: When you use the EXECUTE statement, AlloyDB uses the trained proxy model to process the query locally.
Falls back to the LLM: If the accuracy of the model is low, or if AlloyDB can't find a model, AlloyDB automatically falls back to using the LLM.

Let’s look at the same example of searching for a camera that can handle an underwater depth of 60 meters or deeper using optimized ai.if. Here we train a proxy model using the PREPARE statement and then EXECUTE the statement thereafter.

code_block: <ListValue: [StructValue([('code', "-- Prepare the Optimized Function / Proxy Model\r\nPREPARE waterproof_camera_60m AS\r\nSELECT productid, productname, category, description\r\nFROM products AS p\r\nWHERE\r\n ai.if(\r\n 'Evaluate if the product description indicates that the product is waterproof at depth 60m or deeper. Description:'\r\n || description,\r\n description_embedding);\r\n\r\n-- Run the Proxy Model\r\nEXECUTE waterproof_camera_60m;"), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7f8142bb24f0>)])]>

You see the same products that truly match the criteria of working at a depth of 60 meters - as shown in the screenshot above. Here’s a tabulated version for the first three products, so you can look at the descriptions more closely:

productname	description
Pulsetron Action Camera MZ314	Conquer your next adventure with this camera. Don't let the elements hold you back; dive up to 60 meters deep or withstand rugged trails with its shock-resistant, adventure-ready chassis. Every jump, every turn, every splash is rendered flawlessly smooth with advanced Horizon Lock stabilization, ensuring your footage tells the story with unparalleled fluidity.
Hyperbyte Action Camera LG688	Capture the world in breathtaking detail, even when the action is at its most intense. This camera packs a formidable 1-inch sensor into a remarkably tough, pocket-sized frame. Shoot stunning 5K video and crystal-clear 20MP stills that rival professional equipment. Dive deeper than ever before with robust waterproofing at 60 meters.
Alphasync Action Camera WW897	This formidable, compact camera shrugs off the elements, while the massive 1-inch sensor translates every breathtaking moment into stunning 5K video and crystal-clear 20MP stills. Conquer any environment – from the deepest dive to the highest peak – thanks to its 60 meter waterproofing and revolutionary Horizon Lock, ensuring your footage remains impossibly steady.

See it in action!

Watch how this all comes together in this demo video.

Getting started is easy

Ready to bring unprecedented speed and cost-efficiency to your AI workloads?

New to AlloyDB? Discover AlloyDB with a 30-day free trial.
AI functions quickstart: Enable a few quick prerequisites and start calling functions like ai.if, ai.generate, or ai.analyze_sentiment directly within your SQL queries. Check out these practical examples to begin.
Boost performance and optimize costs: To unlock the biggest performance and cost gains, follow our guide on optimized functions. This is available in preview for ai.if, and will be expanding to more functions soon. For technical insights on this technique, including when it works best and when not, refer to this blog post.
Scale your throughput: Use smart batching to accelerate AI functions (available in preview for ai.if and ai.rank) or array-based functions (generally available for all LLM-based AI functions) to handle bulk prompting smoothly.

Get started with the Claude apps gateway for Google Cloud

Wed, 01 Jul 2026 16:00:00 +0000

Anthropic's agentic coding tool Claude Code has worked with Google Cloud for a while now. An individual developer could easily point CLAUDE_CODE_USE_VERTEX=1 at a Google Cloud (GCP) project, grant the role roles/aiplatform.user, and inference stays inside your Google Cloud perimeter.

That flow works great when it’s just you, or a handful of engineers. But rolling it out across an organization forces you to deal with enterprise friction: you have to manage per-developer cloud credentials, push a managed-settings.json to every laptop over MDM, and not be verified with zero per-developer usage attribution or easily enforceable spend caps.

The Claude apps gateway closes that gap. It is a self-hosted service, shipped with the same claude binary, that sits directly between your local Claude Code clients and Google Cloud. This post breaks down exactly why you should run it and what a secure deployment looks like on Google Cloud.

(Note: If you want to jump straight to the code, the full walkthrough lives in the Claude apps gateway on Google Cloud docs.)

Why run the gateway

Run the gateway to centralize the governance that developers and platform admins otherwise each carry alone such as identity, policy, cost, and routing. Here's what that looks like in practice.

Identity. The /login request routes through your identity provider (IdP ) - Google Workspace or any OIDC/OpenID Connect one - and the gateway swaps the token for a short-lived session. No sensitive information lands on the developer’s laptop — such as service-account keys, API keys, or ANTHROPIC_VERTEX_PROJECT_ID. Onboarding is as simple as adding a user to an IdP group; offboarding by removing them, and their next session refresh fails on the spot.

Policy. Your RBAC (role-based access control) rules live once in gateway.yaml, resolved per group and enforced server-side. The gateway re-checks availableModels on every /v1/messages call, so editing local managed-settings.json changes nothing — and rule updates reach the whole fleet within the hour.

Telemetry. Every claude_code.token.usage metric carries the verified email and groups from the session JWT (signed session token), not the spoofable client-set OTEL_RESOURCE_ATTRIBUTES. The gateway ships them over OTLP/HTTP to a collector you run — Cloud Monitoring, Grafana, Datadog, whatever you use.

Spend limits. Set daily, weekly, or monthly caps per user, group, or org via the admin API; the gateway meters tokens against a Cloud SQL ledger and returns a 429 at the cap. Costs are at list price, so treat them as a runaway-usage guardrail, not a bill reconciliation (committed-use discounts and negotiated rates don't show up).

Routing. Calls go out under a single Cloud Run service identity. Set region: global for Agent Platform's global endpoint, or add a second upstreams: entry to fail over on 5xx/429/timeout in list order. Either way, inference stays in your GCP project — quota, Data Processing Agreement, and billing all unchanged.

How it fits together

A developer's local or deployed claude process sends inference traffic to the gateway over HTTPS. The gateway is a stateless container on Cloud Run as shown below.

The gateway validates its own session bearer — Google Workspace is only contacted at sign-in and token refresh — checks policy, and forwards the request to Agent Platform using the Cloud Run service account. Cloud SQL holds device-code sign-in state and the spend ledger; an OTLP collector receives the attributed metrics.

Setting it up on Google Cloud

The full walkthrough, every gcloud command and the complete gateway.yaml reference, is in the Claude apps gateway on Google Cloud docs. The short version:

Step 1: Provision the GCP foundation
Enable the Agent Platform, Cloud SQL, and Secret Manager APIs; create a claude-gateway service account with roles/aiplatform.user; stand up a small Cloud SQL Postgres database instance for state. The gateway authenticates to Agent Platform as the Cloud Run service identity — you do not create a service-account key. Finally, create a new OAuth client (type Web application) in the Google Cloud console: in this example, the gateway authenticates developers against Google Workspace as an OIDC relying party, and this client is what issues it a client_id and client_secret for that handshake. Those two values feed the oidc: block in the next step. You'll later add the authorized redirect URI once the gateway URL is known.

Step 2: Configure the gateway
Write gateway.yaml pointing at your Google Workspace OIDC client, the Postgres connection string, and Agent Platform as the upstream. Store it in Secret Manager, along with the OIDC client secret, the Postgres URL, and a JWT signing key.

code_block: <ListValue: [StructValue([('code', 'listen:\r\n port: 8080\r\n public_url: https://<your-cloud-run-service-url> # the Cloud Run service URL — with --ingress=internal this resolves only inside your VPC / corporate network\r\noidc:\r\n issuer: https://accounts.google.com # Google Workspace\r\n client_id: <client-id>.apps.googleusercontent.com\r\n client_secret: ${OIDC_CLIENT_SECRET} # from Secret Manager\r\n allowed_email_domains: [yourco.com]\r\n\r\nupstreams:\r\n - provider: vertex\r\n region: us-east5\r\n project_id: <your-project>\r\n auth: {} # ADC via the Cloud Run SA, NO key file'), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7f8154093ca0>)])]>

Then register https://<public_url host>/oauth/callback as an authorized redirect URI on the Google OAuth client — it must match listen.public_url exactly:

Step 3: Deploy to Cloud Run
gcloud run deploy with the service account attached, the Cloud SQL connection on the VPC, and the config mounted from Secret Manager. The container is stateless and scales horizontally behind the Cloud Run load balancer. GKE works equally well if that's already your platform, and only the deployment manifest changes.

code_block: <ListValue: [StructValue([('code', 'gcloud run deploy claude-gateway \\\r\n --service-account="claude-gateway@${PROJECT_ID}.iam.gserviceaccount.com" \\\r\n --set-secrets=/etc/claude/gateway.yaml=gateway-config:latest \\\r\n --ingress=internal \\ # private — developers reach the gateway over the corporate network (VPN/Interconnect into the VPC)\r\n --no-invoker-iam-check # the gateway runs its OWN OIDC; clients carry no GCP token'), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7f8142d1d520>)])]>

Developers connect over the corporate network; you may front the service with an internal Application Load Balancer — see Cloud Run private networking.

Either public or internal, your developers must be able to access whatever URL you configure or you can rely on the default URL from Cloud Run. For the below example we will use https://claude-gateway.example.internal

Step 4: Onboard a developer
Push forceLoginMethod: "gateway" and forceLoginGatewayUrl to developer machines via managed settings. This is how /login knows where to connect, with no manual URL entry. For an org rollout, that's your MDM channel. For a first trial without MDM, the developer can write the file by hand at /Library/Application Support/ClaudeCode/managed-settings.json on macOS (or /etc/claude-code/managed-settings.json on Linux) if they have local admin permissions:

code_block: <ListValue: [StructValue([('code', '{\r\n "forceLoginMethod": "gateway",\r\n "forceLoginGatewayUrl": "https://claude-gateway.example.internal"\r\n}'), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7f8142d1dc70>)])]>

At Claude Code startup, the developer then presses Enter on the pre-filled gateway sign-in screen to confirm the URL.Confirm the device code on the gateway's verification page in the browser, and get redirected to Google Workspace to sign in. After that, the developer completes the device-code flow in the browser against Google Workspace. If setup ends correctly, you will be able to see Cloud Gateway in the terminal view as shown below.

What's next

At this point you should have a better understanding of how to configure and use Claude apps gateway on Google Cloud. Here are some next steps you may want to consider:

Full config reference: every gateway.yaml field is in claude-apps-gateway-config. Per-IdP setup and the GKE track live in claude-apps-gateway-deploy and claude-apps-gateway-on-gcp.
Group-scoped policies: front the gateway with a groups-capable IdP, set groups_claim, and add match: { groups: [...] } policies above the catch-all to give different teams different model lists and tool permissions.

For now, thanks for reading! And if you have any additional questions or feedback, feel free to reach out on socials (Roy Arsan - Linkedin, X and Ivan Nardini - LinkedIn, X)

Happy building!