Google added several capabilities to managed agents in its Gemini API, which run coding and tool-use tasks inside an isolated cloud sandbox. Gemini 3.6 Flash is now the default model with no code changes needed, and developers can attach their own scripts that block, check or log each action the agent takes. Spending caps, scheduled recurring runs, and access from unbilled free tier projects were also added.
What changed
Managed agents required a billed project, ran on the previous Gemini model by default, and offered no way to run custom validation code inside the remote sandbox.
What it unlocks
Running custom scripts before or after each action an agent takes in its cloud sandbox, capping how much a task can spend, and scheduling agent runs on a recurring timetable.
- max_total_tokens caps input, output and thinking combined
- 7-day time-to-live on sandboxes
- hook timeouts of 10 and 15 seconds in the example configuration
What you need to act on it
- a Gemini API key
- free tier projects now supported, no active billing required
Sources