Gemini API Agents Now Default to 3.6 Flash, Add Budget Controls and Hooks

Google has announced a set of updates to Managed Agents in the Gemini API that strengthen cost control, security, and automation capabilities. The antigravity-preview-05-2026 agent now defaults to Gemini 3.6 Flash, a balanced model optimized for reasoning, coding, and tool use. No code changes are required; the next interaction automatically picks up the new default. Developers can also explicitly select other models by passing agent_config.model—choosing Gemini 3.5 Flash (previous generation), Gemini 3.5 Flash-Lite (lowest latency and cost), or pinning to a model of choice.

Environment hooks represent a major capability expansion. Developers can now add a .agents/hooks.json configuration file to define custom scripts that execute before or after every tool call inside the sandbox. For example, a security-gate hook can run a Python script before code_execution or write_file calls; if the script returns {"decision": "deny", "reason": "..."}, the tool call is skipped. An auto-format hook can run after every tool to enforce code styling. Hooks support regular expression matchers (e.g., "code_execution|write_file" or catch-all "*") and can be command-type (running local scripts with a timeout) or HTTP-type (POSTing to an external endpoint). Offdeal, an AI-native investment bank, uses post_tool_execution hooks to validate company logos in financial decks—fetching candidates, enforcing pixel-level quality checks, verifying each logo with Gemini vision, and publishing a manifest of approved files inside the remote sandbox, enabling banker-ready presentations without external validation pipelines.

Cost control features address the risk of runaway spending in multi-turn autonomous loops. Developers can pass max_total_tokens inside agent_config to cap total consumption (input + output + thinking). When the agent reaches the limit, execution safely pauses and the interaction returns status: "incomplete". The environment state is preserved, so you can resume by passing previous_interaction_id with a fresh budget. Scheduled execution with triggers lets developers bind an agent, environment, prompt, and cron schedule into a persistent resource that fires without manual intervention; each run reuses the same sandbox, so files persist across executions. The Environments API lets you list, inspect, and delete sandbox sessions from code, recover environment IDs after disconnect, and clean up sandboxes when your pipeline finishes instead of waiting for the 7-day TTL.

Free tier availability rounds out the release: Managed Agents are now available on free tier projects, allowing developers to experiment with agentic workflows using an API key from a project without active billing. Together, these updates transform Managed Agents into cost-controlled, scheduled workers operating autonomously inside real development environments without breaking budgets or requiring external orchestration.

Managed Agents in the Gemini API have evolved from a foundational capability into a production-grade system addressing real operational concerns: cost predictability, security validation, and autonomous scheduling. The introduction of environment hooks directly responds to a concrete use case—Offdeal, an AI-native investment bank, needed to verify financial deck assets (company logos) at pixel-level quality inside the sandbox rather than externally. With hooks, Offdeal's Archie AI analyst can now fetch candidates, enforce quality checks, and use Gemini vision to validate each logo without leaving the managed agent sandbox.

The shift to Gemini 3.6 Flash as the default reflects both model maturity and user preference for speed and cost in agentic workflows. Budget controls (max_total_tokens) acknowledge that multi-turn autonomous loops can accumulate tokens unpredictably; the ability to pause safely and resume with a fresh budget preserves debugging and iteration patterns without surprise bills. Free tier access expands the constituency: teams can now prototype agentic workflows on API keys without enabling billing, lowering friction for early adoption.

Scheduled triggers and the Environments API round out the package, positioning Managed Agents as a self-contained worker system that can run recurring tasks, persist state across runs, and be managed programmatically—all within an isolated cloud sandbox, with no external orchestration required.

No comments yet. Be the first to share your thoughts!