Google has updated its Gemini API managed agents to run on Gemini 3.6 Flash by default, introducing tool execution hooks, token budget controls, scheduled triggers, and free tier access.
Google announced the changes on July 28, 2026, confirming that the antigravity-preview-05-2026 agent picks up Gemini 3.6 Flash automatically without requiring code modifications. Developers can also explicitly select models by passing configuration parameters, including the previous-generation Gemini 3.5 Flash or the lower-cost Gemini 3.5 Flash-Lite.
Environment hooks and validation
The update adds environment hooks that run custom scripts inside isolated cloud sandboxes before or after tool calls. By adding a .agents/hooks.json file to the environment, systems can intercept pre_tool_execution or post_tool_execution events.
The configuration supports regular expression matchers to target specific tools or catch all actions. If a pre-execution script denies an action, the runtime skips the tool call and injects the rejection reason into the model context. Hooks support local command execution as well as HTTP handlers that post to external endpoints.
OffDeal, an AI-native investment bank, implemented post-execution hooks to validate company logos generated by its AI analyst, Archie. Founder and Chief Technology Officer Alston Lin stated that hooks allow their pipeline to run pixel-level quality checks and logo verification directly within the remote sandbox as soon as company lists are generated.
Budget controls and scheduling
Google introduced cost controls to manage autonomous execution loops. Developers can pass a max_total_tokens parameter to cap combined input, output, and thinking tokens. When an agent reaches this ceiling, execution pauses and returns an incomplete status while preserving environment state, allowing tasks to resume under a new token budget using the prior interaction identifier.
The platform also added scheduled execution through triggers, which bind an agent, environment, prompt, and cron schedule into a persistent automated resource that reuses sandboxes across runs. Developers can manage these sessions through the Environments API, which provides programmatic listing, inspection, and manual deletion of sandboxes ahead of their standard seven-day time-to-live limit.
