Gemini API managed agents: background tasks, remote MCP

Google is adding background execution, remote Model Context Protocol server integration, custom function calling and credential refresh to Managed Agents in the Gemini API, announced Jul 07, 2026. The updates extend the Gemini Interactions API managed agents that run reasoning, code execution, package installation, file management and web information inside an isolated cloud sandbox.

What did Google add and how do the new features work?

The release adds four capabilities: long-running background execution, direct connections to remote MCP servers, custom function calling alongside built-in sandbox tools, and network credential refresh. Background execution uses a single-flag API call: pass "background: true" to run interactions asynchronously; the API immediately returns an ID clients can poll, stream progress from, or use to reconnect while the agent finishes remotely. Remote MCP server integration lets developers pass an mcp_server tool at interaction time so managed agents can communicate with private endpoints from the sandbox. Custom tools run locally by transitioning an interaction to "requires_action" while built-in sandbox tools run automatically on the server. Network credential refresh works by passing your existing environment_id with a new network configuration on the next interaction; the new rules replace the old ones immediately and the sandbox retains its filesystem state, installed packages and cloned repositories.

The Interactions API model remains a single-endpoint flow: you call one endpoint and Gemini handles reasoning and execution inside the cloud sandbox. For AI coding agents the announcement includes a concrete install instruction for the Interactions API skill: npx skills add google-gemini/gemini-skills --skill gemini-interactions-api.

How will these pieces fit into an agent workflow?

Background execution turns managed agents into asynchronous workers: clients send an interaction with "background: true", receive an interaction ID, and then poll or stream status while the agent performs long-running tasks without holding an HTTP connection open. Remote tools mix with built-in sandbox capabilities; developers can include Google Search or code execution alongside an mcp_server tool to reach internal APIs without custom proxy middleware. The API’s step-matching behavior determines where logic runs: built-in sandbox tools execute automatically on the server, while custom functions cause the interaction to require client-side action, labeled "requires_action".

Credential lifecycle is handled by environment-level updates. Passing the same environment_id with a new network configuration rotates or refreshes tokens and keys immediately; the sandbox preserves filesystem state and existing installations so long-running agent state is not lost during a credential update.

Why it matters

Managed agents already wrapped reasoning, code execution and file handling inside an isolated cloud sandbox. Making those agents asynchronous with background execution addresses the fragility of holding HTTP connections open for long tasks, and direct MCP integration reduces the need for bespoke proxy middleware to reach private data. Allowing custom functions to hand control back to clients while built-in tools run server-side gives developers a clear separation between safe, server-run operations and local business logic execution. Credential refresh without wiping sandbox state lowers operational friction for agents that must maintain environment continuity.

What to watch

Check the background execution guide and the managed agents quickstart for implementation details and streaming patterns. Also watch for any published best practices around remote MCP server use and the security rules developers should follow when mixing remote tools with built-in sandbox capabilities.