Gemini's agent API gets hooks and spend

Google added a batch of production controls to Gemini API Managed Agents on July 28. The service now defaults to Gemini 3.6 Flash, with 3.5 Flash and 3.5 Flash-Lite selectable, and no code change needed for existing integrations. Environment hooks let a developer run custom scripts before or after any tool call, configured via .agents/hooks.json with pre_tool_execution and post_tool_execution events, regex matchers, and HTTP-based handlers, aimed at security gating, linting, and auditing inside the sandbox.

Cost control gets a hard ceiling: a max_total_tokens parameter caps input, output, and thinking-token consumption combined, and an interaction that hits the limit pauses safely with status: "incomplete" rather than failing outright, resumable later via previous_interaction_id. Scheduled triggers bind an agent to a cron-style schedule for recurring autonomous runs, with sandbox state persisting across executions, and a new Environments API lets a developer list, inspect, and delete sandbox sessions directly instead of waiting on the default seven-day cleanup. Managed agents are now available on free-tier projects with no billing required.

What it means for operators

Every major agent platform has been converging on the same three gaps: unattended-job cost limits, tool-call-level auditing, and a scheduling primitive that doesn’t require a separate cron job wired into your own infra. Google shipping all three in one update, on top of ADK 2.0’s move toward explicit graph-based control flow, is the kind of concrete spend-governance feature the agent-budgeting playbook has been arguing every team needs before it lets an agent run unattended. If you’ve been holding off on scheduled Gemini agents because a runaway loop had no real stop condition, max_total_tokens plus the pause-and-resume behavior is the specific feature that removes that objection.