Google Deepmind just gave its Managed Agents in the Gemini API a serious upgrade. The big one? Background Execution lets your agents keep working even when no HTTP connection is open — think of it as giving them a to-do list they can chip away at offline. Another killer feature is native support for remote MCP servers, which means agents can ping your internal databases or APIs directly without extra middleware. Plus, you can now mix custom functions with the built-in sandbox tools, giving you more control over what agents can do. And there’s a clever credential refresh that lets you swap tokens mid-conversation without blowing up the sandbox environment — a huge time-saver for long-running tasks. All of these land in the Gemini Interactions API, with code snippets in JavaScript, Python, and cURL. Why it matters: These features quietly shift AI agents from toys you babysit to workers you can trust to handle multi-step workflows autonomously.