Google launches Gemini API Managed Agents

Google launched Managed Agents in the Gemini API on May 19, adding a Google-hosted runtime for developers building autonomous software agents. The company said the feature is in public preview and lets developers start an agent with a single API call. Google’s launch materials describe the service as a way to run reasoning, tool use, code execution and web access inside an isolated Linux sandbox. A May 24 DEV Community guide by Pulkit Govrani walked through setup steps and examples after the I/O rollout. ### What did Google actually release? Google said on May 19 that the Gemini API “now supports managed agents,” extending infrastructure the company had previously used for its own agent products. In the launch post, Google DeepMind staff members Ali Çevik and Philipp Schmid said developers can spin up an agent that reasons, uses tools and executes code in an isolated, ephemeral Linux environment. (blog.google) The Gemini API release notes said the launch put Managed Agents into public preview and introduced the general-purpose Antigravity agent, listed as `antigravity-preview-05-2026`. Google said the agent can autonomously plan, reason, write and execute code, manage files and browse the web inside its sandbox container. ### What does “managed” mean here? Google said the managed part is the runtime layer. (blog.google) In its launch post, the company said production-grade agents previously required developers to manage infrastructure, scaffolding and isolated sandboxes as they scaled. Managed Agents, Google said, abstracts that work so developers can focus on product behavior and agent logic. (ai.google.dev) The Agents overview page says a single API call provisions a Linux sandbox where the agent can reason, execute code, manage files and browse the web autonomously. Google also says developers can extend the default agent with their own instructions, skills and data. ### How are developers supposed to customize one? Google said custom agents can be defined through markdown files including `AGENTS.md` and `SKILL.md`. (blog.google) The launch post said developers can register those files as a managed agent instead of writing more complex orchestration code. The same post said Managed Agents are powered by the Antigravity agent, built on Gemini 3.5 Flash and available through the Interactions API and Google AI Studio. (ai.google.dev) Google’s documentation also points developers to AI Studio for prototyping agents without writing code. ### What are the constraints and safeguards? Google said every managed agent runs in an OS-level isolated sandbox, but the company also warned users to review outputs before relying on them in sensitive workflows. (blog.google) The Agents overview says outbound network access is unrestricted by default, though developers can restrict or disable it with an allowlist. Google also said credentials can be injected securely and are not exposed inside the sandbox, while urging developers to use least-privilege credentials and short-lived tokens. (blog.google) The documentation says a single interaction can consume roughly 100,000 to 3 million tokens, and environment compute is not billed during the preview. ### Where does this fit in Google’s broader I/O push? (ai.google.dev) Google said at I/O 2026 that it was expanding Antigravity, its agent-focused development platform, alongside the Gemini 3.5 model family. In a separate developer-highlights post, the company said Managed Agents give developers access to the same agent harness and infrastructure behind Google’s own products. (ai.google.dev) The DEV Community guide published on May 24 by Pulkit Govrani described the product as a fast route from prompt to deployed agent and focused on setup and examples. Google’s official materials say the next entry points are the Agents overview, the Quickstart and new custom templates in Google AI Studio, while enterprise support for managed agents on Gemini Enterprise Agent Platform is in private preview. (blog.google) (developers.googleblog.com)