Google Cloud Expands AI Model Choice and Tools for Multi-Agent Systems
Editor’s note: For a monthly roundup of Google Cloud announcements, resources, events, and learning opportunities, check back here.
October: More models, more ways to deploy
Google Cloud’s October announcements are about giving enterprises more choices for building AI systems, especially systems that coordinate multiple agents. The breadth is notable: new model access sits alongside tools for government and robotics, plus practical guidance for developers. Whether that adds up to safer production deployments will depend on how these offerings work in practice; the announcements alone don’t settle that question.
Top announcements
- Gemini 4 Argon comes to Google Cloud. Google says Argon is designed to keep context through long, multistep work, with domain expertise to support complex software builds, financial and legal tasks, and cybersecurity tools such as CodeMender.
- Anthropic-managed Claude deployments become generally available. Teams can now choose this deployment option on Google Cloud, alongside the existing Google-managed option, and consolidate AI workloads on the platform.
- Google Cloud adds OpenAI models. Google says it is now the only hyperscale cloud provider with all three leading closed-source LLMs, as well as more than 200 first-party, third-party, and open models on one enterprise AI platform.
- Claude Opus 5.5 and Claude Sonnet 5.5 arrive in Google Antigravity. Google positions Opus for complex work and Sonnet as a faster, lower-cost option, extending model choice for Google Cloud customers.
- Gemini for government services enters preview. The new industry suite includes government-focused skills, connections to agency systems, partner agents, and underlying governance.
- Gemini Robotics Embodied Reasoning 2 becomes available on Google Cloud. Organizations can combine Google DeepMind’s embodied reasoning model with Google Cloud security, multi-agent orchestration, and lifecycle management to simulate, assess, and deploy physical AI.
Thought leadership: the editor’s pick
How startups are advancing robotics with Google AI looks at work by startups following the launch of Gemini Robotics Embodied Reasoning 2 on the Gemini platform and AI Studio. The model brings Google DeepMind’s multimodal reasoning and long-horizon task planning for robotics to Google Cloud. Darren Mowry, Google’s VP of Global Startups and Investor Ecosystem, introduces the examples.
News you can use
- A developer’s guide to MCP apps in Gemini walks through implementation. The case for visual, multistep interfaces is straightforward: text-based enterprise assistants can make each adjustment another slow chat exchange.
- Four evaluation engineering lessons from building agent plugins shares practices developers can apply to their own builds.
- Who judges the judge? Three ways to audit LLM evaluators examines barriers to strong evaluations and offers a three-pillar method for improving quality.
For more on what Google Cloud customers are building, see the monthly roundup, Cool stuff customers built.
Start building with $300 in free credit to try Google Cloud AI and ML.
September
September’s coverage ranged from model updates and security tools to new implementation guides. Google highlighted Gemini 3.8 Live with Live Avatar and Mantis, its open-source bug-finding and fixing framework. The month also brought the announcement of Gemini 4 Argon, which Google describes as a frontier model for software engineering, enterprise knowledge work in areas such as law and finance, and cybersecurity defense. More details are available here.
Top announcements
- Gemini 3.8 Live with Live Avatar is generally available in Gemini Enterprise, following the announcements of Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking.
- Getting started with Mantis, Google’s open-source bug-finding and fixing harness. Google says Mantis is based on its internal approach to finding and fixing vulnerabilities at machine speed, using scalable, context-aware repository analysis.
Thought leadership: the editor’s pick
Formula E brings real-time AI to the edge follows the team’s attempt to set records at the Goodwood Hillclimb. Its setup used Gemini Enterprise, Gemma, and a Pixel phone mounted on the car’s dashboard to explore AI and edge computing.
News you can use
- Best practices for customizing Gemini models with reinforcement learning covers the RL training loop, when to use reinforcement learning fine-tuning, and how to get more from it.
- The Google Cloud CLI remote MCP server is in preview. Built around gcloud and bq, the BigQuery command-line tool, it gives AI agents access to command-line operations for managing Google Cloud infrastructure and running advanced BigQuery workflows.
- A guide to graph workflows in ADK uses a refund workflow to explain fan-out and fan-in, deterministic and agent routers, human review pauses, parallel workers, and dynamic orchestration. It also considers when to use a fixed graph and when Python should determine what runs next.
For a customer roundup, see Cool stuff customers built.
August
August’s updates focused on making business AI more practical, including controlling its costs and adapting it to specialized work. Google’s examples centered on financial services and legal teams, alongside billing and spending controls for agent workloads.
Top announcements
- Flexible billing and cost controls for agents add expanded billing options and cost-management tools across Gemini Enterprise and developer tools, including Google Antigravity in Gemini Enterprise and Android Studio.
- Gemini Enterprise for Financial Services brings Google’s agentic AI into capital markets and corporate banking workflows.
- Gemini Enterprise for Legal offers an integrated, governed environment configured for deployment across law firms and corporate legal departments.
- Google Antigravity expands to enterprise customers through eligible Gemini Enterprise app subscriptions, with administrative and spending controls included.
Thought leadership: the editor’s pick
July: moving agents toward production
The July updates point to a shift in how companies are using AI: experiments are turning into production systems, and the hard questions now involve security, architecture and control. Google Cloud paired platform announcements with practical guides, including 13 demos and 20 diagnostic questions for engineering teams planning their systems.Platform updates
- What’s new in Gemini Enterprise Agent Platform covers capabilities now available to everyone, including Agent Runtime and Agent Identity.
- Now in preview: Find and fix software vulnerabilities with CodeMender makes the case for automated code remediation as AI speeds up attacks. The documentation has more on the preview.
- Solve harder problems with AlphaEvolve, now available to everyone on Google Cloud introduces a Gemini-based code optimization and discovery agent for difficult algorithmic problems in business and research.
Trust and control
Delegating work to an AI agent means trusting it to act within boundaries. The productivity case may be clear, but losing control remains a real concern, especially when agents encounter unusual data. Google Cloud’s July editorial picks focus on transparency, predictable behavior and setting those limits.- How leaders can scale AI by trading control for trust (Q&A) features Michael Gerstenhaber, VP of Product Management for Gemini Enterprise, on why safe boundaries matter.
- What makes an AI agent trustworthy looks at the role of company context, with guidance from Prajakta Damle, Senior Director of Product Management.
Guides to put to work
- Automate your agent development lifecycle using any coding agent explains how Agents CLI skills can support the stages of agent development from within a coding agent.
- Why AI apps fail in production (And how Google solved it) says only 5% of AI prototypes reach production; the other 95% stall during validation. It asks how teams can make that transition with more confidence.
- 13 hands-on demos to build on Gemini Enterprise Agent Platform offers a set of starting points for teams exploring the platform.
June: building and securing AI systems
June’s updates covered ways to build, scale and secure AI applications while keeping them under human control. One demonstration took that work into retail: at Cannes, Google Cloud showed a virtual shopping experience designed to make product discovery more engaging.Platform and product news
- Introducing the Open Knowledge Format presents OKF, an open specification that turns the LLM-wiki pattern into a portable format. It is designed to let people and AI agents share metadata, context and curated knowledge across vendors.
- Collaboration with Apple on its expanded Private Cloud Compute (PCC) systems describes a partnership grounded in privacy, using Google Cloud’s Confidential Computing portfolio and Titanium security architecture.
- Claude Fable 5: Available on Google Cloud announces general availability for Anthropic’s latest frontier model, part of Google Cloud’s effort to bring current models to its Agent Platform.
- Cloud Atelier: How Gemini Enterprise is helping restyle the retail playbook describes the destination-based virtual shopping experience presented at Cannes, where browsing becomes a product discovery experience.
Security lessons and leadership
Google Cloud’s security coverage makes a pointed argument: AI is speeding up threats, so defenses need to respond at comparable speed. The company says it has embedded specialized agents in its software development lifecycle to create automated code guardrails, and that it plans to make those protections more widely available.- How Google Cloud Security uses AI internally explains the company’s move toward a proactive, more autonomous security posture.
- The 4 lessons that guided AI Threat Defense introduces Chris Betz as Google Cloud’s new CISO and shares four lessons from applying AI to defense.
Practical reading
- 5 lessons from red teaming AI applications outlines Mandiant’s risk-based approach to secure AI development, centered on the Good AI Assessment (GAIA) Top 10 and its report, Secure Development of Generative AI Applications: A Proactive Approach.
- How to unlock true ROI in software development – a deep dive into the latest DORA research draws on the DORA: ROI of AI-assisted software development report to help teams assess costs, benefits and early adoption challenges.
- Agent Factory Recap: 100X engineering with AI agents in Google Antigravity 2.0 brings together Shir Meir Lador, Head of AI Engineering at Google Cloud Developer Relations, and Google agentic engineer Rody Davis to discuss agent-first platforms, code review and the use of “skills.”
May: spending, reliability and hands-on learning
May’s selected reading takes up a practical tension in AI adoption: teams need to manage costs while making models more reliable and turning prototypes into working systems.- Tokenomics: Why smart teams spend more on AI, on purpose features Eric Lam, Head of Value, Delta, Google Cloud Consulting, on how disciplined organizations are moving beyond reactive concern over AI bills and adopting a deliberate approach to token spending.
- Meet the researcher fighting AI hallucinations at Google Cloud profiles Cyrus, a Google senior research scientist whose recent work focuses on hallucinations: cases where a model lacks the right facts but confidently invents an answer.
- What sports cars can teach us about optimizing AI spend features Mike Clark, Director of Product Management for Gemini Enterprise Agent Platform. His central point: more tokens don’t automatically produce better AI, so teams need to balance model capacity with efficiency and returns.
Guides for building
- 10 questions every startup should answer before moving to production with their AI prototype focuses on the prototype-to-production transition. Each question includes a short runnable snippet. Data layers, RAG architecture, CI/CD and network design sit outside its scope.
- Your chance to start building AI agents from the absolute basics introduces Agent Valley, a free, five-week live learning series where participants build agent systems in a small, low-poly virtual world.
Google Cloud’s I/O announcements center on agents
The thread running through Google Cloud’s I/O news is clear: AI models are being packaged with tools that can take action, build software and support security teams. That shift matters more than any single launch, though the announcements alone don’t show how well these products will perform in day-to-day work.
For a broader account of the I/O news, Google has published a guide for Google Cloud customers and a separate roundup of news for startups.
Top announcements
On security, Google introduced Google AI Threat Defense, an always-on autonomous platform that uses AI for cybersecurity. Google describes it as a comprehensive solution; the announcement’s practical test will be how it helps defenders keep pace with attackers. Read the announcement.
The rest of the lineup puts agents at the center of Google’s model and developer strategy:
- Gemini 3.5 is a new model family designed to pair advanced intelligence with the ability to act. Gemini 3.5 Flash is the first model in the family.
- Gemini Omni is built for broader world understanding, multimodal work and editing. Google says it can generate different kinds of output from different kinds of input, beginning with video.
- Google Antigravity gains expanded capabilities and an integration with Agent Platform, bringing agentic development to organizations.
- Gemini Spark, for Gemini Enterprise and Workspace customers, is a personal AI agent intended to work around the clock and take actions under the user’s direction.
- Google Workspace adds Google Pics for image generation and editing, along with voice features in Gmail, Docs and Keep.
- Managed Agents API on Agent Platform lets developers build and run custom agents in secure, Google-hosted environments connected to Agent Platform.
- CodeMender, an AI security agent available through Agent Platform, can help identify and fix code vulnerabilities.
Google also made Nano Banana 2 and Nano Banana Pro generally available through Gemini Enterprise Agent Platform. The company says organizations are already using the models. More on availability.
Security across clouds
One editorial pick looks at what the Google and Wiz combination could mean for multicloud security. Vinod D’Souza, director of Google’s Office of the CISO, recounts an RSA Conference conversation with Anthony Belfiore, Wiz’s chief strategy officer. Their argument is that combining Wiz’s cloud telemetry with Google’s AI and quantum research can help CISOs address threats, including the adversarial use of AI, as organizations adopt agentic systems. The announcement doesn’t spell out how that approach will work in every multicloud environment, but it puts shared security data and AI capabilities at the heart of the pitch. Read the discussion.
From announcements to deployment
If you’re building agents on Google Cloud, the next question is how the pieces fit together. Google’s guide to developing agents after I/O covers Gemini Enterprise Agent Platform, the new developer tools and the available choices for building.
For teams trying to move beyond prototypes, a five-part series collects architecture patterns and practices for putting agents into production. There’s also guidance on preparing public-sector security programs for AI, covering environments from industrial control systems to long-established municipal databases.
Google says it will continue sharing monthly updates on its AI announcements, news and practices. Its customer roundup, “Cool stuff customers built,” offers examples of what organizations are making with Google Cloud.
April: Google Cloud Next
That focus on agents carried into Google Cloud Next, held in Las Vegas on April 22. Google announced products ranging from Gemini Enterprise Agent Platform to its eighth-generation TPUs, and added collaborative features to the Gemini Enterprise app. Projects, for example, lets people work alongside agents and colleagues.
Those who missed the livestream can read the Day 1 recap. Google says it has counted more than 1,300 customer examples of AI use. The number signals broad experimentation; it doesn’t, by itself, tell us how many deployments have reached production or delivered lasting results.