Google Updates Gemini API Managed Agents With Gemini 3.6 Flash and Sandbox Hooks

Google expands Managed Agents in the Gemini API

Google DeepMind has updated Managed Agents in the Gemini API with Gemini 3.6 Flash as the default model, new sandbox-level environment hooks, budget controls, scheduled triggers and free-tier access.

The changes were outlined in a Google Blog post by Philipp Schmid, a member of the technical staff at Google DeepMind, and product manager Mariano Cocirio. They build on an earlier Managed Agents release that added background tasks and integration with remote Model Context Protocol (MCP) servers.

Gemini 3.6 Flash becomes the default

According to Google, the antigravity-preview-05-2026 agent now runs on Gemini 3.6 Flash by default. The company says existing users do not need to make code changes to receive the new default.

Managed Agents are available through the Gemini Interactions API. Google describes the API as a way to coordinate reasoning, code execution, package installation, file management and web retrieval through a single API call, with those operations taking place in an isolated cloud sandbox.

Google's developer documentation also describes the Interactions API as providing managed-agent support, server-side state, observable execution steps and background execution. Those capabilities position the service as infrastructure for longer-running software tasks rather than only a prompt-and-response model endpoint.

Hooks add controls around tool use

The most notable operational addition is environment hooks. Google says these hooks can be used inside the sandbox to block, lint or audit tool calls.

That creates a point of control around actions an agent attempts to take while completing a task. For example, a development team could use hooks to apply its own checks before a tool call proceeds, or to retain an audit trail of activity occurring in the execution environment. Google did not describe the hooks as replacing broader application-level security controls; instead, the announcement presents them as controls within the managed sandbox.

Budgets, schedules and access changes

Google also announced budget controls and scheduled triggers for Managed Agents. The company’s post does not detail the specific configuration options in the supplied announcement, but the features indicate that developers can set constraints and initiate work on a schedule rather than only through an immediate request.

In addition, Google said Managed Agents are available on a free tier. That expands access to the feature set for developers evaluating the Interactions API or prototyping workflows that require code execution, file operations or web retrieval.

For developers using JavaScript or TypeScript, Google’s examples use the @google/genai SDK. The company directs Python and cURL users to its Antigravity agent documentation.

What the update means

The update combines a new default model with more direct supervision of tool activity and additional controls for running tasks over time. For teams adopting Managed Agents, the practical questions will be how environment hooks fit into their existing review and authorization practices, and how budget and scheduling settings behave across different workloads.

Google’s announcement makes clear that the Gemini API is continuing to package model reasoning and execution tooling into a managed service, while adding controls intended to make those automated operations easier to govern.