Google's Gemini Managed Agents Get a Quieter Upgrade: Faster Default Model, Guardrails, and a Free Tier
Gemini 3.6 Flash is now the default for managed agents, and new environment hooks let developers block or audit what the AI does inside its sandbox.
Key points
- Google's Gemini API managed agents now default to Gemini 3.6 Flash, with no code changes required for existing users.
- New environment hooks let developers run custom scripts before or after any tool call the agent makes inside its cloud sandbox.
- A new
max_total_tokenssetting caps how much an agent can spend on a task before it safely pauses. - Managed agents are now available on free tier Google AI projects, so developers can try them without active billing.
- Scheduled triggers let agents run on a recurring timetable, reusing the same sandbox between runs.
Google has quietly upgraded the tools it offers developers who build AI agents, software that can carry out multi-step tasks like editing files or running code on its own. The changes, announced on the DeepMind blog, focus less on raw model power and more on keeping agents cheap and predictable.
When we covered Gemini 3.6 Flash's debut on 21 July 2026, it was pitched as a faster, cheaper foundation for autonomous software. Now it's the default inside the agent runtime itself.
What is a managed agent, in plain words?
A managed agent is an AI worker that Google runs inside an isolated cloud sandbox, a small walled-off computer. You give it a task in one API call. It reasons, writes code, installs packages, then returns a result.
Developers use these to build automated code reviewers, AI research tools, or analyst features inside financial apps. The user sees a helpful button. All the complexity runs out of sight.
What changed with the default model?
The antigravity-preview-05-2026 agent now runs on Gemini 3.6 Flash by default, replacing Gemini 3.5 Flash. Existing code picks it up automatically on the next call.
Google describes 3.6 Flash as balanced for reasoning and coding. Developers who want to cut costs can pin the agent to Gemini 3.5 Flash-Lite, the lowest-latency option Google currently offers in that family.
| Model | Role |
|---|---|
| Gemini 3.6 Flash | New default, balanced |
| Gemini 3.5 Flash | Previous generation |
| Gemini 3.5 Flash-Lite | Lowest cost and latency |
Why do the new hooks matter?
Environment hooks let a developer run their own script before or after every action the agent takes. A company can block risky tool calls or audit what the AI did, without trusting the model to police itself.
A developer drops a .agents/hooks.json file into the environment. The runtime fires their scripts on pre_tool_execution or post_tool_execution events. If a pre-execution script returns a deny decision, the tool call is skipped and the reason is passed back into the model's context so it knows to try something else.
Alston Lin, founder and CTO of OffDeal, an AI-native investment bank, said his team uses post-execution hooks to verify company logos inside investment banking decks. Each logo gets checked for the right company, aspect ratio, transparent background, and contrast before it lands in a slide. "Before agent hooks, we couldn't do this on Gemini's managed agents," Lin said, in comments reported by Google. "The sandbox is remote, so our validation code had nowhere to run."
For ordinary users, this is the difference between an AI feature that occasionally embarrasses a company and one that has a real safety net.
What about runaway costs?
Agents that loop autonomously can burn through tokens, the small chunks of text an AI charges by, faster than anyone notices. Google's new max_total_tokens cap, passed in when creating an interaction, stops that.
When the agent hits the limit, it pauses with a status of incomplete and preserves its environment. Developers can resume it with a fresh budget by passing the previous interaction ID. Scheduled triggers let an agent run on a cron schedule too, reusing its sandbox so files persist between runs.
Who can try it now?
Managed agents are now available on free tier projects. Developers can experiment using an API key from a Google AI project with no active billing required. There's also a new Environments API for listing and deleting sandbox sessions from code, rather than waiting for the default seven-day timeout.
None of this is a headline-grabbing model launch. It's the plumbing that decides whether the agents built on Gemini feel reliable or reckless. Watch whether hooks get adopted widely enough to become a de facto compliance layer for enterprise AI tools: that's the real test.
