📰 Key Highlights

Google announces new features for Gemini API’s Managed Agents, directly responding to developer feedback, with the goal of making agents more suitable for production deployment.

At the architectural core, Managed Agents runs through a single API endpoint, with Gemini automatically handling reasoning, code execution, package installation, file management, and web information retrieval in an isolated cloud sandbox — developers don’t need to manage the underlying infrastructure themselves.

This update brings four major new capabilities. First, background async execution: in the past, long-running tasks required keeping an HTTP long connection alive, which was easy to drop. Now, just pass background: true, and the API immediately returns a task ID. The client can poll progress anytime or reconnect without blocking the main application. Second, remote MCP server integration: developers can connect Managed Agents directly to remote Model Context Protocol servers to access private databases or internal APIs, no more building a custom agent middleware layer, and you can mix them with built-in tools like Google Search and code execution. Third, custom function calling: built-in tools run automatically on the server side, while custom functions trigger a requires_action state, letting the client run business logic locally. The two work together through a step-matching mechanism. Fourth, dynamic network credential refresh: when access tokens or short-lived API keys expire, just bring the original environment_id and pass in the new network config. The old rules are replaced immediately, while the sandbox’s file system, installed packages, and cloned repository state are all preserved intact.

💬 JudyAI Lab Perspective

Google’s latest update to Managed Agents directly addresses three pain points developers hit in production: long-task connection drops, building custom middleware just to access private data, and losing the entire sandbox state when credentials expire.

What’s worth paying attention to isn’t just the feature list — it’s a shift in design philosophy: push the complex infrastructure integration down to the platform layer, so developers can focus on business logic. Background async execution means long connections are no longer the bottleneck. Remote MCP integration lets agents hook into private databases directly, no need to write an extra adapter proxy. And the division of labor in custom function calling draws a clear line between “the platform runs this automatically” and “you run this business logic locally.” We’re seeing this direction become mainstream — agent infrastructure is looking more and more like utilities, and developers are increasingly not having to wire up the entire underlying pipeline themselves.

If you’re designing an agent architecture, you can test every decision with one question: “Does this part have to be managed by me, or can the platform handle it?” The places where you can’t articulate a clear reason are usually the engineering choices worth thinking about carefully.

📅 Source Information

🔗 Further Reading