Google Cloud launches Agent Registry and governance tools for enterprise AI agents

Google Cloud is trying to turn AI agents into something enterprises can actually operate. Not just demo. Not just “we built a bot.” The gap has been pretty obvious — companies can spin up agents, but once there are dozens of them, plus tools, connectors, and permissions, the whole thing starts to look ungovernable. At Google Cloud Next ’26, held April 22 to 24, Google answered that with Gemini Enterprise Agent Platform, plus a new Agent Registry and a broader set of governance and observability tools. (cloud.google.com) ### What actually launched? Google folded Vertex AI’s model and agent-building pieces into a broader product called Gemini Enterprise Agent Platform. The pitch is simple — one platform to build, scale, govern, and optimize agents, instead of stitching those jobs together across separate products. Google described it as the evolution of Vertex AI, with new layers for integration, DevOps, orchestration, and security. (cloud.google.com) ### What is Agent Registry? Agent Registry is the control catalog inside that platform. It gives companies a central place to store, discover, and manage AI agents, tools, and Model Context Protocol servers across the organization. That sounds boring, but it is the boring part that matters. If agents are going to call tools, share context, and hand tasks to(cloud.google.com)ng twice. (docs.cloud.google.com) ### Why does MCP show up here? Because Google is betting that enterprise agents will not live inside one vendor’s box. MCP — Model Context Protocol — is becoming one of the ways agents connect to tools and data sources. By making Registry a catalog for MCP servers as well as agents, Google is basically saying the important object to manage is not just the model. It is the whole working system around the model — tools, endpoints, permissions, and shared context. (docs.cloud.google.com) ### What do the governance tools actually do? They cover the stuff enterprises always ask for once the pilot ends — identity, IAM policies, sharing controls, safety settings, and operational oversight. Google’s docs frame this as governance plus inventory, with monitoring for performance, utilization, and system health. In other words, the platform is trying to answer two different questions at once: “What is this agent allowed to do?” and “What is this agent actually doing?” (docs.cloud.google.com) ### How does observability fit in? Google also added observability features that use OpenTelemetry-style signals, including dashboards, topology views, and traces for agent behavior. That matters because multi-agent systems fail in weird ways. One agent calls another, which calls a tool, which hits a policy wall, and now nobody knows where the task broke. Traces give platform teams a way to debug that chain instead of treating the agent as a black box. (docs.cloud.google.com) ### Why is this different from last year? A year ago, the center of gravity was still model access and chatbot deployment. Now Google is packaging the agent layer as infrastructure. Even the Gemini Enterprise app is being positioned as something built on top of the platform, inheriting governance, security, and identity controls by default. That is a shift from “here is an AI assistant” to “here is an operating environment for many assistants and automations.” (cloud.google.com) ### So what is Google really selling? Basically, a control plane for enterprise AI agents. The flashy part is autonomous software. The sticky part is registry, policy, and monitoring. If companies really are moving from one-off copilots to networks of agents, the winners may be the vendors that make those networks legible and governable. That is the lane Google is trying to own here. (cloud.google.com)