Google adds background tasks and remote MCP to Gemini Managed Agents

Why it matters

Google is moving Gemini's agent layer from chat-style calls toward hosted, stateful workers. That puts pricing, permissions and reliability at the center of the developer decision.

Logan Kilpatrick (@OfficialLoganK), Google's product lead for Google AI Studio and the Gemini API, said on X on July 7th that Google is rolling out new Managed Agents features in the Gemini API, including background tasks, remote MCP, function calling, credential refresh and access through the free tier.

The update, detailed the same day in a Google blog post, is a practical one: Google is trying to turn Managed Agents from a demo-friendly sandbox into something developers can leave running against real workloads. Kilpatrick wrote that the goal is to "reduce the cost, friction, and complexity" of putting agents into production, and said Google has seen "thousands of customers" using the API. That customer count is Google's claim, not an independently disclosed usage metric.

Kilpatrick is the right messenger for this particular push. His Google author page identifies him as product lead for Google AI Studio and the Gemini API at Google DeepMind, where he leads the team focused on helping developers build with Gemini, Veo, Imagen and related tools. The product bet is familiar from his work on AI Studio: Google wants developers to stay inside its API path as agent workflows move from chat completions to longer-running tasks that need state, tools and credentials.

Managed Agents sit on top of the Gemini Interactions API, which Google describes as the main interface for Gemini models and agents. In the Agents Overview, Google says one call provisions a Linux sandbox where the agent can reason, execute code, manage files and browse the web. The default managed agent, Antigravity Agent, is powered by Gemini 3.5 Flash and runs through the Interactions API in Google AI Studio and the Gemini API.

The most important addition is background execution. Instead of keeping an HTTP connection open while an agent researches, runs code or moves through a multi-step plan, developers can set background: true. Google's background execution documentation says the API returns an interaction ID immediately, then lets applications poll status, stream progress or reconnect later while the server keeps running the job. That matters for production applications because long agent runs commonly exceed normal request timeouts.

Remote MCP support is the second production-facing change. Google says developers can register remote Model Context Protocol servers as tools, using the mcp_server tool type in the request. The Antigravity docs specify that remote MCP servers must run over streamable HTTP, with fields for name, URL, optional headers and allowed tools. That gives Managed Agents a standard path into private databases, internal services and third-party tool servers without forcing every developer to write a custom proxy layer.

Function calling fills a different gap. Google's docs say custom functions can be added alongside built-in sandbox tools. Built-in tools run on Google's side; custom business logic can push the interaction into requires_action, where the developer's client executes the function and returns the result. For teams already using function calling to keep sensitive operations inside their own systems, that split is the control point.

Credential refresh is the operational fix buried inside the announcement. Google's environment documentation says developers can pass an existing environment_id with a new network configuration to refresh tokens or rotate short-lived keys. The new network rules replace the old ones, while the sandbox keeps files, installed packages and cloned repositories. Google also says credentials injected through header transformations are handled by an egress proxy and are not exposed inside the sandbox as files or environment variables.

The free-tier line in Kilpatrick's post is meant to lower the trial barrier, but pricing remains a constraint for serious use. Google's pricing page says Managed Agents are charged based on underlying Gemini model inference, including intermediate input and reasoning tokens generated during agent loops, plus tool usage where applicable. Google says environment compute is not billed during the preview period. The Antigravity docs warn that agent interactions can accumulate large token counts because a single request can trigger repeated reasoning, tool use, code execution and file operations.

Managed Agents also remain a preview product. Google's docs say developers should review an agent's actions and outputs before relying on them in sensitive workflows, especially when the agent modifies data or interacts with external systems. The new features do not remove that risk. They make the product easier to wire into real applications, which raises the importance of permissions, allowlists and short-lived credentials.

The timing fits Google's broader Gemini developer strategy. In a separate July developer post, Google said the Interactions API is the default for Google AI Studio, the Gemini API and its documentation, while the older generateContent API remains supported. The direction is clear: frontier agent capabilities are being routed through the stateful Interactions API, where Google can package models, tools, sandboxes, files and execution state behind one interface.

Kilpatrick's reply to Maximiliano Moxarella (@maxi_moxa) captured the pressure behind the launch: the goal, he wrote, is to let developers run these agents "without it breaking the bank," adding that smaller models hopefully help. That is the unresolved commercial question for Google's hosted-agent push. The infrastructure is getting closer to a cloud worker abstraction for agents. The adoption curve will depend on whether developers can predict what those workers cost before they let them run.