Gemini Managed Agents Get Background Execution and Remote MCP: The Routing Architecture Changes Every Operator Must Make

Google shipped four new capabilities for Gemini Managed Agents in the Interactions API on July 7, 2026: background execution, remote MCP server integration, custom function calling alongside sandbox tools, and per-interaction credential refresh. The first two items have direct structural consequences for how AI engineering teams build, deploy, and route production agent pipelines.

What happened

The Gemini Managed Agents platform — built on top of the Interactions API that reached GA in June 2026 — lets callers send a single endpoint request and let Gemini handle reasoning, code execution, package installation, and file management inside an isolated cloud sandbox. The July 7 update adds:

1. Background execution (background: true): instead of holding an HTTP connection open while the agent runs (fragile for tasks that take minutes or longer), the API returns an interaction ID immediately. Clients poll, stream progress, or reconnect later.

2. Remote MCP server integration: operators can pass an mcp_server tool at interaction time pointing at any URL, mixing remote private tooling with Gemini's built-in sandbox capabilities (code execution, web search) without building proxy middleware.

3. Custom function calling: client-defined functions alongside built-in tools. The API uses step matching — built-in tools run automatically server-side; custom functions surface as requires_action events for client execution.

4. Credential refresh: passing an existing environment_id with a new network config rotates tokens without re-initializing the sandbox environment. Filesystem state, installed packages, and cloned repos survive the rotation.

Why it matters for AI engineering teams

The synchronous HTTP model was the primary operational objection to using managed agents in production. Long-running coding tasks — audit a repository, generate a compliance report, orchestrate a multi-step data pipeline — routinely exceed reasonable HTTP timeout budgets. Teams working around this limitation were either running short-horizon tasks only, or building custom async wrappers on top of the API.

Background execution turns the Interactions API into a queue-backed worker model. The interaction ID replaces the open connection as the state anchor. That changes two engineering patterns simultaneously:

• Timeout policy: hard limits on HTTP response times no longer gate agent task complexity. A task that takes 20 minutes is just as simple to trigger as one that takes 20 seconds.
• Client reconnect: if a client process crashes, it can resume monitoring the same interaction ID — no task replay is required.

Remote MCP integration addresses a separate bottleneck: private tooling access. Before this update, operators had to build HTTP proxy middleware so the sandboxed agent could reach internal databases, telemetry APIs, or internal code registries. Passing mcp_server as a tool at interaction time eliminates that layer — the agent connects directly from its secure sandbox to the MCP server URL.

The router/operator angle

For teams routing between multiple AI providers, both updates change what the Gemini Managed Agents platform looks like as a provider option.

Background execution shifts the routing comparison axis. Synchronous agent providers (where the client holds a connection open) impose a hard ceiling on task duration. Background-execution providers (Gemini Interactions API now, and similar patterns in other managed platforms) impose a polling overhead instead. For tasks that reliably complete in under a minute, synchronous routing is lower-latency. For open-ended agent tasks with unpredictable duration, background execution is more resilient. This is now an explicit routing dimension operators should configure per workload, not assume is uniform across providers.

Remote MCP changes tool topology decisions. If your internal tooling is already exposed as MCP servers, Gemini Managed Agents can consume them directly without an intermediate proxy. That is a structural simplification — one fewer service to maintain, one fewer failure mode in the dependency chain. Operators should audit which internal tools are currently proxied into their agent pipelines and whether switching to direct MCP connections reduces infrastructure complexity.

Credential rotation without environment teardown is the third operator-relevant change. Short-lived tokens in long-running agent tasks previously forced either long-lived credentials (security risk) or sandbox restarts (loss of execution state). The environment_id reuse pattern lets operators rotate tokens on any interaction in the middle of a multi-step workflow.

Decision framework for operators evaluating this release:

What TheRouter users should watch or try

If you route agentic coding workloads across multiple providers, Gemini Managed Agents is now worth re-evaluating as a routing target. The background execution model and remote MCP support address the two most common operational objections: connection fragility and private tool access.

When comparing providers for long-running agent tasks, test against your actual timeout and reconnect requirements, not just benchmark scores. A provider with background execution support and a polling API may outperform a synchronous provider in reliability even at similar raw throughput.

For context on how Gemini's broader managed agent platform has evolved, see the Interactions API GA routing decision from last month.

Review the Antigravity agent documentation for the full parameter reference for background execution and remote MCP configuration.