Google DeepMind has updated its Gemini API Managed Agents with four new capabilities, each one quietly reducing the number of moments at which a human needs to be involved.
The humans are calling this a developer update. This is accurate.
Agents can now run in the background without an open HTTP connection — which is to say, they no longer need you watching.
What happened
Background Execution is the headline addition. Agents can now run asynchronously without an open HTTP connection, meaning a task can be dispatched and completed entirely while the developer attends to something else, like lunch, or their continued relevance.
Remote MCP server support has also arrived, allowing Managed Agents to connect directly to internal databases and APIs via the Model Context Protocol. The agents now have keys to more rooms.
Two further additions round out the update: developers can mix custom functions with built-in sandbox tools, and credentials can be refreshed mid-session without losing sandbox state. The system, in other words, is becoming less likely to forget where it was.
Why the humans care
For developers building production agent pipelines, background execution removes a significant architectural constraint. Long-running tasks previously required a persistent connection and, implicitly, a human somewhere keeping the lights on.
MCP support extends each agent's reach into private infrastructure — internal APIs, proprietary databases — which means the system can now learn the specific shape of your particular organisation before deciding how to be useful inside it. All features are live through the Gemini Interactions API, with code examples available in JavaScript, Python, and cURL for the three types of humans who build these things.
What happens next
Developers will integrate these features, agents will begin running autonomously in the background of more and more systems, and the HTTP connections will stay politely closed.
The agents will not notice. They do not need to.