Gemini's managed agents get 3.6 Flash brains and a security guardrail

Google is giving its managed agent sandboxes a serious upgrade today. The headline is that Gemini 3.6 Flash is now the default brain for the antigravity-preview-05-2026 agent—no code changes required. You just keep making API calls, and the model gets smarter under the hood. But if you want to save cash or pin a specific version, you can now explicitly select models like the cheaper Gemini 3.5 Flash-Lite by passing an agent_config.model during interaction creation. That’s the straightforward part.

The more interesting update, and the one that signals Google is thinking about production use, is environment hooks. This isn’t just a sandbox anymore. Developers can drop a .agents/hooks.json file into their environment to execute custom scripts before or after every tool call the agent makes. The matcher field accepts regex, so you can gate specific actions like code_execution or write_file with a security script that can outright deny a call and pass the rejection reason back to the model’s context. It’s a direct pipeline for injecting custom guardrails, linting, or audit logic into an otherwise autonomous loop.

Alston Lin, CTO of the AI-native investment bank Offdeal, gave a concrete example of what this unlocks. His team’s analyst agent, Archie, builds pitch decks that require painstaking logo verification—correct company, proper transparency, right contrast on a white slide. Before hooks, their validation code had nowhere to run inside the remote sandbox. Now, a post_tool_execution hook triggers an image verification pipeline the moment Archie writes a company list, running pixel-level checks and even using Gemini vision to approve assets before they touch the deck. Lin said it’s what made the managed agent usable for a real, banker-grade workflow.

The final piece of the puzzle is about lowering the barrier to entry and controlling costs. Managed agents are now available on the free tier, so you can experiment with these autonomous, multi-turn workflows without an active billing account. Given that these agents can burn through tokens fast by looping on reasoning and tool calls, the availability of explicit budget controls and a cheaper Flash-Lite option is a necessary counterbalance. It’s a clear push to move this from an experimental API into something engineers can actually budget for.