Google is turning agent runtime infrastructure into a Gemini API feature, not just a model prompt.
- Google says Managed Agents in the Gemini API can spin up an agent that reasons, uses tools, and executes code in an isolated, ephemeral Linux environment.
- The first path uses Google’s Antigravity agent, built on Gemini 3.5 Flash, through the Interactions API and Google AI Studio.
- The rollout is preview, while enterprise support through the Gemini Enterprise Agent Platform is private preview, so this is not a broad general-availability claim.
Google has moved one layer deeper into the AI-agent stack. In a May 19 announcement, the company said the Gemini API now supports Managed Agents: cloud-hosted agent sessions that can provision a remote Linux environment, use tools, execute code, manage files, and browse the web.
That matters because the hard part of production agents is not only choosing a model. Teams also need orchestration, tool access, file state, sandboxing, and a way to resume work without rebuilding the whole environment around every request. Google’s pitch is that developers can call the Gemini API and get more of that runtime layer handled for them.
What Google Announced
Google says Managed Agents can be started with a single call. The launch centers on the Antigravity agent, which Google says runs in a secure cloud sandbox and is powered by Gemini 3.5 Flash. In that environment, the agent can reason, plan, call tools, execute code, manage files, and browse the web to fetch and process live data.
The file-state piece is important for builders. Google says each interaction either creates or receives an environment, and follow-up calls can resume the same session with files and state intact. That means an agent can keep working across steps instead of treating every request like a fresh chat.
Google also says developers can customize agents with their own instructions and skills. Instead of writing all orchestration logic as application code, teams can define agent behavior in markdown files such as AGENTS.md and SKILL.md, then register those definitions as a managed agent.
Why This Is More Than Another API Feature
For a normal model API call, the application usually owns the surrounding machinery: tool routing, sandbox setup, temporary files, execution boundaries, and session state. Managed Agents point to a different product shape. The model provider is no longer only selling inference; it is also packaging the harness where agent work happens.
That could lower the barrier for teams that want agents to perform multi-step technical work. A product team testing a research assistant, coding helper, data-processing agent, or internal operations bot may not want to maintain its own sandbox fleet before proving the workflow. Google is positioning Gemini Managed Agents as a way to start closer to the behavior layer while Google handles more infrastructure.
It also fits a broader shift in AI competition. Model quality still matters, but developer platforms increasingly compete on the surrounding runtime: how agents connect to tools, how safely they execute code, how state persists, and how easily teams can version instructions and skills.
The Sandbox Is The Key Caveat
Google’s announcement repeatedly emphasizes isolation. The agent runs in an isolated, ephemeral Linux environment. That does not remove all security, data, or governance questions, but it is a meaningful product signal: agent platforms are being designed around the assumption that agents will touch files, execute code, call tools, and browse live web data.
For developers, the upside is practical. Sandboxes can make experiments faster because the agent has a place to work. For enterprises, the harder questions remain around access controls, auditability, policy enforcement, data retention, and which workloads are safe to delegate. Google says enterprise support for managed agents on the Gemini Enterprise Agent Platform is in private preview, not public general availability.
What Builders Should Watch Next
The near-term question is availability. Google says Managed Agents in the Gemini API are rolling out in preview, with Google AI Studio templates also planned for getting started. That makes this a builder-facing launch, but not yet proof that the feature is mature across production use cases.
The second question is portability. Defining agents through files like AGENTS.md and SKILL.md gives teams a versionable surface, but the actual runtime behavior still depends on Google’s managed infrastructure. Teams evaluating the feature should separate portable agent instructions from platform-specific execution behavior.
The third question is whether managed sandboxes become a default layer for serious agent products. If developers adopt this pattern, the agent market may move from “which model answers best?” toward “which platform gives agents the safest and fastest place to do work?”
Google’s Managed Agents launch makes the Gemini API look less like a plain model endpoint and more like a cloud agent runtime. The announcement is strongest when read narrowly: Google is offering preview access to managed agent infrastructure with isolated Linux sandboxes, Antigravity agent support, session state, and file-based customization. What is not verified is broad general availability, production maturity across customer workloads, or enterprise access beyond private preview.
