How Gemini Managed Agents Works under the Hood

Gemini Managed Agents work by spinning up a full execution environment behind a single API call. When interactions.create() is called, the API boots an isolated Linux VM (Ubuntu, 4 vCPU, 16 GB RAM), loads skills, and enters a reasoning loop where Gemini 3.5 Flash picks tools, executes code in the sandbox, reads outputs, and repeats until the task is complete. Environments can persist across interactions via an environment_id. The setup scales from ad-hoc calls to production managed agents saved with client.agents.create(), with each invocation forking a fresh sandbox from the base environment.