Google announces new capabilities for Managed Agents in the Gemini API, including 3.6 Flash and hooks, enabling developers to build more reliable, production-ready agents.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Google DeepMind has rolled out several updates for Managed Agents on the Gemini API: the default model is upgraded to Gemini 3.6 Flash, new environment hooks enable interception, auditing, and formatting of tool calls within the sandbox, and budget controls, scheduled triggers, and free tier access are introduced.
Google DeepMind announced that the default model for Managed Agents has been upgraded to Gemini 3.6 Flash. According to the official blog, the antigravity-preview-05-2026 agent now runs Gemini 3.6 Flash by default, and users do not need any code changes—the next interaction will automatically use the new model.
Developers can also explicitly select other models: via the agent_config.model parameter, they can specify Gemini 3.5 Flash or Gemini 3.5 Flash-Lite. Gemini 3.5 Flash-Lite is described as the model with the lowest latency and cost in the Gemini 3.5 series. This flexibility allows users to balance performance and cost.
The newly introduced environment hooks allow developers to run custom scripts before and after each tool call within the agent sandbox. By adding a .agents/hooks.json file to the environment, the runtime executes the corresponding handler when pre_tool_execution or post_tool_execution events are triggered.
Hook configuration supports regular expression matching for tool names. For example, a security-gate group can run a gate.py script before code_execution or write_file calls; if the script returns {"decision": "deny", "reason": "..."}, the tool call is skipped and the reason is passed to the model context. Another example is the auto-format group, which runs auto_lint.py after each tool execution to enforce code style.
Hooks also support HTTP-type handlers that can POST directly to external endpoints. The official documentation provides complete HTTP hook definitions and failure handling semantics.
The blog cites the case of AI-native investment bank Offdeal: its AI analyst Archie needs to ensure that company logos are displayed correctly in slides—including correct company, size, aspect ratio, transparent background, and high contrast on white slides. Offdeal uses post_tool_execution hooks to trigger an automatic image verification pipeline within the remote sandbox.
Offdeal stated that before agent hooks existed, verification code had nowhere to run because the sandbox was remote; hooks enable the verification pipeline to execute inside the sandbox. This demonstrates the practical value of environment hooks in production-grade verification workflows.
The blog also mentions new budget controls, scheduled triggers, and free tier access, but does not provide specific details. These features build on previously released background tasks and remote MCP server integration.
Managed Agents coordinate reasoning, code execution, package installation, file management, and web retrieval through the Gemini Interactions API, all within isolated cloud sandboxes.
This article's information primarily comes from the official Google blog, which is a first-party source and thus highly credible. However, specific details on budget controls, scheduled triggers, and free tier are not elaborated in the source, so related descriptions are general. The Offdeal use case is cited from the blog and is a third-party claim, not independently verified.
Google has significantly enhanced the productivity and controllability of Managed Agents through the default model upgrade and environment hooks, particularly suitable for enterprise scenarios requiring custom verification and auditing within the sandbox.
Primary report
Primary source