Google's Gemini Managed Agents Get Async Tasks and Direct MCP Access

Google is rolling out a series of infrastructure upgrades for its Gemini API managed agents — the AI agents that run inside Google's isolated sandbox environments. The changes tackle three major pain points that have been plaguing developers trying to deploy production-ready agents: long-running tasks dropping due to HTTP timeouts, the need to build custom proxy middleware just to connect to private data sources, and the frustration of having your entire sandbox — installed packages, cloned repos, the works — reset every time you refresh an API key.

Four new capabilities directly address these issues:

1. Async background execution: Pass background: true to run tasks asynchronously in the cloud. The API immediately returns a task ID, and clients can poll for status or reconnect to receive progress without keeping a persistent connection open.
2. Direct remote MCP connectivity: Agents can now talk directly to remote Model Context Protocol (MCP) servers — no more proxy intermediaries needed to securely access private databases or internal APIs.
3. Hybrid function calls: Sandbox tools run server-side, but custom local functions can flag themselves with requires_action, letting the client execute specific business logic locally while the rest runs in the sandbox.
4. Credential refresh without sandbox reset: When rotating short-lived API keys or credentials, developers can use environment_id to overwrite configs without losing the sandbox's file system, cloned repositories, or installed dependencies.

Managed agents are already available in the API's free tier, and developers can get started with the @google/genai SDK.