Google Cloud launches Gemini agent platform

AI agent platforms are starting to look less like model menus and more like operating systems. That is the real news here. On April 22 at Google Cloud Next ’26, Google launched Gemini Enterprise Agent Platform — basically a rebuilt, expanded Vertex AI aimed at companies that want agents to do actual work inside business systems, not just answer prompts. The pitch is simple: stop stitching together models, runtimes, security, and orchestration by hand, and use one managed stack instead. (cloud.google.com) ### What actually launched? Google says Gemini Enterprise Agent Platform is the new developer platform for building, scaling, governing, and optimizing agents. It folds Vertex AI’s model selection, tuning, and app-building pieces into a broader system with new agent integration, orchestration, DevOps, and security features. Google is also making this the path fo(cloud.google.com)rate standalone product. (cloud.google.com) ### Why rename Vertex AI at all? Because the problem changed. Two years ago, enterprises mostly wanted access to a strong model and some tooling around it. Now they want fleets of agents that can call tools, hold state, remember context, and operate across apps with guardrails. Google is trying to say the center of gravity has moved from “model development” to “agent operations.” The rename is branding, sure — but it is also product positioning. (cloud.google.com) ### What sits inside the platform? A lot. Google’s docs break it into four pillars — build, scale, govern, optimize. On the build side, there is low-code Agent Studio, the code-first Agent Development Kit, Model Garden, Agent Garden, RAG Engine, and Vector Search. On the scale side, there is Agent Runtime, sessions, persistent Memory Bank, and sandboxed code exe(cloud.google.com) what Google thinks enterprises are really buying: not intelligence alone, but controlled execution. (docs.cloud.google.com) ### Why does the 200-plus model number matter? Because Google is not pretending one model family will win every workflow. The platform offers access to more than 200 models through Model Garden, including Google’s Gemini 3.1 Pro, Gemini 3.1 Flash Image, Lyria 3, open models like Gemma 4, and third-party options including Anthropic Claude Opus, Sonnet, and Haiku. That is a very deliberate (docs.cloud.google.com)e’s model. (cloud.google.com) ### What is the technical catch? Agents are harder than chatbots because they persist. Google’s new runtime supports long-running agents that can maintain state for days, with Memory Bank for long-term context. That is useful, but it also creates the hard enterprise problems — permissions, audit trails, policy enforcement, prompt-injection defense, and agent spr(cloud.google.com) of optional add-ons. (cloud.google.com) ### Where does this fit in Google’s bigger push? It sits in the middle of Google’s “agentic enterprise” strategy. Google said nearly 75% of Google Cloud customers are already using its AI products, and 330 customers processed more than 1 trillion tokens each over the past year. The company is pairing the new platform with TPUs, data products, Workspace integrati(cloud.google.com)oyee interface. (blog.google) ### So what changed for buyers? The buying decision is shifting upward. Enterprises are no longer just comparing model benchmarks. They are choosing an agent control plane — the layer that decides how models, tools, memory, identity, and policy fit together. If Google wins that layer, it matters less whether a given workflow runs on Gemini, Gemma, or Claude. (cloud.google.com)## Bottom line This launch is Google telling enterprises that the model race is becoming infrastructure. The flashy part is “200-plus models.” The important part is that Google wants to be the place where companies govern thousands of agents without losing control. (cloud.google.com)