Use Gemini Managed Agents for long-running MCP work without babysitting timeouts
Google expanded Managed Agents in the Gemini API with background execution, remote MCP server integration, custom functions, and credential refresh. The practical shift is simple: agent runs can last longer, call more tools, and survive more real workflow plumbing without forcing developers to keep a client connection open.
Google confirmed on July 7, 2026 that Managed Agents in the Gemini API now support background execution, remote MCP server integration, custom function calling, and credential refresh. Confidence level: confirmed. The update matters for builders who want agent runs to continue after a client disconnects, call external tools through MCP, and keep authentication usable across longer interactions.
What changed
Google added four production-facing capabilities to Managed Agents in the Gemini API: background execution for asynchronous work, remote MCP server integration, custom function calling, and network credential refresh across interactions. The announcement is dated July 7, 2026.
Managed Agents run through the Gemini Interactions API. Google describes the model as a single endpoint where Gemini handles reasoning, code execution, package installation, file management, and web information inside an isolated cloud sandbox.
The practical change is not a new frontier model. It is more runtime plumbing around agents: longer jobs, external tools, custom APIs, and credentials that can stay usable across multi-step workflows.
Why this is early
The first signal came from Google's own developer announcement and supporting docs, not from a leak. Independent aggregation picked it up the same day, but the strongest evidence remains Google's blog post and Gemini API documentation.
This is still an early builder story because Managed Agents and the Antigravity agent surface are preview-oriented. The useful question is not whether the feature exists, but whether its current limits fit your workflow, security model, and cost tolerance.
Key takeaways
- Background execution helps long-running agent tasks avoid client timeout failures.
- Remote MCP lets a Gemini agent connect to external tools and services through server configuration.
- Custom functions give teams a more controlled way to connect business APIs than asking an agent to browse around.
- Credential refresh reduces friction for workflows that span multiple interactions.
- The feature belongs in a controlled pilot first, especially when external systems or credentials are involved.
| Capability | Best fit | Access/status | Cost/status | Caveat |
|---|---|---|---|---|
| Background execution | Long research, coding, and file-generation tasks | Gemini API Managed Agents | Check current Gemini API billing | Requires polling or retrieval logic |
| Remote MCP | Tool access through MCP servers | Interactions API tools config | Depends on model/API use and tools | Server name and transport limits matter |
| Custom function calling | Calling internal APIs with schemas | Gemini API function calling | Standard API/tool costs may apply | You still own validation and authorization |
| Credential refresh | Longer authenticated workflows | Managed agent environments | Depends on configuration | Scope credentials narrowly |
Availability and access
Google says Managed Agents are built on the Gemini Interactions API, and the docs show examples for Python, JavaScript, and REST. The Antigravity agent is the managed agent used in the current examples.
Users should verify account access in Google AI Studio or through their Gemini API project before planning a rollout. Pricing, quota, preview availability, and region support can vary by account and product surface.
Practical LinkLoot angle
This update is useful for teams that already have agent prototypes but keep losing time to runtime glue: polling loops, tool adapters, sandbox setup, and auth refresh. Gemini Managed Agents move more of that into Google's hosted agent layer.
The tradeoff is control. Hosted sandboxes are easier to start, but you need a clear policy for network access, credential scope, audit logs, generated files, and failed runs. Treat the first pilot like infrastructure work, not a model demo.
For adjacent tooling, keep LinkLoot's guide to AI agent tools nearby when comparing Gemini Managed Agents with self-hosted runners, coding-agent CLIs, and MCP-first stacks.
What to verify before you act
- Confirm Managed Agents are enabled for your Google AI project and region.
- Check the exact Antigravity agent, Interactions API, and model IDs available to your account.
- Review remote MCP transport and naming constraints before wiring existing MCP servers.
- Inspect how credential refresh stores, scopes, and rotates access to external systems.
- Run a small failure-mode test: timeout, bad tool response, revoked credential, and malformed function output.
Source check
Confirmed by:
- Google announced the Managed Agents expansion on July 7, 2026.
- Gemini API docs describe managed-agent customization, remote environments, tools, remote MCP, and function calling.
Early signal / context:
- Digg tracked same-day public context around the update and linked the Google announcement and docs.
LinkLoot will treat pricing changes, broader GA status, new model support, or major security-limit updates as update triggers rather than assuming preview behavior will stay fixed.
Google added background execution, remote MCP integration, custom function calling, and credential refresh to Managed Agents in the Gemini API.