Google has expanded Managed Agents in the Gemini API with controls that make the hosted agent runtime look more like scheduled infrastructure.
The July 28 update says Managed Agents now default to gemini-3.6-flash, support environment hooks around tool calls, run on free-tier projects, accept total-token budgets, support scheduled triggers, and expose an Environments API for inspecting or deleting sandbox sessions.
That is a meaningful follow-up to Google’s earlier Managed Agents launch. The first release gave developers a hosted Gemini agent with code execution, file management, package installation, web retrieval, background tasks, and remote MCP integrations. This update adds more of the surrounding operational surface: how teams gate tool calls, control cost, repeat jobs, and clean up agent environments.
Hooks put policy inside the sandbox loop
The most important control is environment hooks. Google says developers can add an .agents/hooks.json file so scripts run before or after tool calls inside the remote sandbox.
That gives teams a concrete place to block, lint, audit, or validate actions before the agent continues. A security gate can intercept code execution or file writes. A formatting hook can run after a tool finishes. HTTP hooks can send events to external services.
Hooks do not decide the whole trust model. They do give the application a programmable checkpoint between the model’s intended action and the sandbox’s execution surface.
Google’s example customer uses post-tool hooks to verify generated company logos for banker-ready decks before publishing an approved manifest. That example is Google-reported, but the design pattern is broader: use hooks for deterministic checks the model should not grade by itself.
Budgets and triggers make agents repeatable
Autonomous agent runs can consume tokens unpredictably because they plan, inspect, retry, and revise. Google now lets developers pass max_total_tokens inside agent_config. When the agent hits the limit, execution pauses with an incomplete status while preserving the environment so the caller can continue with a fresh budget.
Scheduled triggers move the same runtime toward recurring work. A trigger binds an agent, environment, prompt, and cron schedule so the task can run without a manual API call. The Environments API then gives teams a way to list, inspect, and delete sandboxes instead of waiting for automatic expiry.
Those additions shift Managed Agents from a demo-friendly API to a platform component. The product question becomes less “can the model use tools?” and more “can the organization run this worker safely every day?”
