Gemini API Gains Remote MCP and Asynchronous Background Tasks

Google updated Managed Agents in the Gemini API to support long-running background tasks and direct integration with remote Model Context Protocol (MCP) servers. The release shifts the API’s focus from synchronous chat generation to production-grade autonomous runtimes. These capabilities rely on the newly recommended Interactions API, which replaces legacy endpoints for workflows requiring multi-step processing and state retention.

Asynchronous Background Execution

Developers can now run interactions asynchronously by passing background: true in the API call. This architectural shift prevents timeout failures during complex, multi-step operations like repository cloning, extensive codebase analysis, or CI/CD pipeline monitoring.

Instead of holding an HTTP connection open while the model works, the API immediately returns an Interaction ID. Client applications use this identifier to poll for status updates via client.interactions.get(id). By decoupling the network request from the task duration, multi-agent systems can orchestrate longer tasks without implementing custom queuing middleware or risking socket timeouts.

Remote MCP and Custom Sandboxing

The update introduces native support for remote servers using the Model Context Protocol. Managed agents can now securely access private data silos and internal APIs directly from Google’s isolated cloud sandbox.

Developers configure this by passing an mcp_server object at interaction time. This allows the agent to mix built-in Google sandbox tools, like web search or Python execution, with proprietary enterprise functions. Prior to this release, integrating internal observability data or proprietary databases required custom proxy layers to handle the routing and serialization.

This integration is accompanied by a new hybrid execution model that allows custom function calling inside the sandbox environment. Agents can execute dynamically generated code within the managed Linux environment while simultaneously triggering external API calls defined by the developer.

State Management and Credential Refresh

To support extended execution cycles, Google introduced a credential refresh mechanism across interactions. Long-running autonomous operations frequently span across standard authentication token lifecycles, causing mid-task failures. The new mechanism maintains session state automatically, ensuring the agent retains the necessary authorization context when executing multi-turn sequences.

These updates are primarily targeted at the Antigravity agent, a general-purpose environment powered by Gemini 3.5 Flash. The framework’s focus on auditability and state management via tracking a previous_interaction_id addresses core enterprise requirements for predictable, stateful AI agents.

If you are migrating existing generateContent workflows to the Interactions API, deploy the deployment skill via npx skills add google-gemini/gemini-skills --skill gemini-interactions-api. Transitioning to the interaction-based polling model allows your agents to handle latency-heavy external network calls without dropping the session state.