Google added four capabilities to Managed Agents in the Gemini API on July 7: agents that can now run in the background, direct connections to remote MCP servers, support for custom function calls, and the ability to refresh credentials mid-session. The update, posted on blog.google by Google DeepMind’s Philipp Schmid and Mariano Cocirio, targets developers who have complained that agent connections break on long tasks and that reaching private tools requires custom middleware. Google frames it as production hardening, not a new product.
The core pitch is abstraction. Developers hit one endpoint on the Interactions API. Behind it, Gemini reasons, executes code, installs packages, manages files, and pulls information from the web, all inside an isolated cloud sandbox. That is a different architecture than the harness pattern most agent builders use today, where a developer’s own server holds the loop open, manages tool calls, and stitches together a sandbox from separate services. Google is betting developers would rather rent that orchestration than build it themselves.
Background execution is the most consequential change. Passing background: true on an interaction returns a task ID right away, so an app never has to keep a socket open while the agent grinds through a job. Developers can check that ID later, watch progress stream in, or step away and reconnect once the sandbox finishes running. Google’s own documentation describes holding a connection open for long-running work as “fragile,” a tacit admission that the earlier, synchronous version of managed agents was not built for jobs that take minutes or hours, like cloning a repository and triaging every open TODO inside it.
MCP, Anthropic’s protocol for tool calling, is the connective tissue behind the second change. Rather than build custom proxy middleware to reach a private database or an internal API, developers can now point a managed agent straight at a remote MCP server. That tool sits alongside Gemini’s built-in code execution and search inside the same sandboxed request, so a single agent call can pull from a company’s internal telemetry system and a public index at once. Google building first-class support for a protocol Anthropic authored is a signal that MCP has become the default wire format for agent tool access, regardless of which lab ships the underlying model.
Custom function calling and credential refresh round out the release. The API now separates tools that run automatically inside the sandbox from custom functions, which pause the interaction in a requires_action state so the client executes local business logic instead of exposing it to Google’s servers. Credential refresh lets a developer rotate an expiring API key or access token against an existing environment_id, and the sandbox keeps its filesystem state, installed packages, or cloned repositories intact through the swap.
This puts Gemini’s managed agents in direct competition with OpenAI’s Assistants and Agents SDK and Anthropic’s own Claude Agent SDK, both of which already abstract sandboxed execution behind a hosted API. None of the three labs has published independent benchmarks comparing agent reliability or task completion rates across platforms, so for now the comparison runs on feature checklists rather than measured performance.
The tradeoff is dependency. A team that builds around Gemini’s managed sandbox, its filesystem persistence, its network rules, its credential model, is writing against Google’s environment rather than a portable one. Teams currently maintaining a self-built agent harness should weigh that lock-in against the real engineering cost of the connection-management and tool-proxy problems Google just solved for them, before locking in a vendor for 2026 agent infrastructure.
Google detailed the update in a July 7, 2026 post on blog.google, written by Google DeepMind’s Philipp Schmid and Mariano Cocirio.
