Managed Agents API

The traditional approach to building an agent requires assembling components yourself: choose a model, connect tool definitions, provision a code execution sandbox (an isolated environment where code runs safely), wire up memory, and handle the orchestration loop. A Managed Agents API collapses that into a single API call. You describe what the agent should do; the provider handles infrastructure.

Google's version, shipped at I/O 2026, lets developers spin up an agent that reasons, uses tools, and executes code inside an isolated Linux environment with one call to the Gemini API. Sessions are persistent, meaning the agent's files and state survive after you close the tab and can be resumed later. This is a meaningful difference from stateless chat APIs, where context resets every session.

The pattern is distinct from traditional serverless inference (where you just run a model call) because it provisions a full agent runtime, not just a model endpoint. It sits alongside tools like E2B and Modal in the category of managed execution environments, but at a higher abstraction level. For builders, the trade-off is the usual one: less control and potential vendor lock-in in exchange for dramatically reduced infrastructure work.