Agent development is no longer about wrangling servers and assembling sandboxes — it's about defining a config file and making a cloud call. At I/O 2026, Google launched the preview of Gemini API Managed Agents, opening up its underlying agent infrastructure to developers. With one API call, you can instantly spin up a temporary, isolated Linux environment hosted by Google, running an Antigravity agent powered by the new mid-tier model Gemini 3.5 Flash.
The managed approach eliminates the need for complex orchestration code and environment setup. Instead of writing lengthy glue code, developers just create two versioned Markdown files — AGENTS.md and SKILL.md — to define the agent's system persona and specific skills, then register it as a custom managed agent.
This ephemeral cloud Linux environment gives agents a surprising amount of agency. They can perform multi-step reasoning, plan tasks, call tools, execute code, read and write files, and even fetch real-time data from the web. Each interaction automatically saves or restores the sandbox's state, so subsequent API calls seamlessly resume the session with all files and runtime intact.
The approach is strikingly similar to the Claude Managed Agents Anthropic released back in April. Both large-model companies are converging on the same vision: evolving from pure API model providers into container cloud platforms that manage secure sandboxes and session state for agents.
Managed Agents are now available in preview on the Gemini API. Developers can quickly get started using custom templates in the Google AI Studio Playground. Google has also introduced a new Gemini API agent skill specifically for coding agents, enabling them to build and manage their own custom agents. For enterprise customers, the Gemini Enterprise Agent Platform is also entering private preview.