Google DeepMind Updates Gemini API Managed Agents With Hooks and Budget Controls

Google expands controls for Gemini API Managed Agents

Google DeepMind has updated Managed Agents in the Gemini API with Gemini 3.6 Flash as the default model, environment hooks for tool-call oversight, and controls for token budgets and scheduled execution.

The changes apply to Managed Agents in the Gemini Interactions API, which Google describes as a service that coordinates reasoning, code execution, package installation, file management and web retrieval through a single API call. Those activities run in an isolated cloud sandbox, according to the company’s announcement.

Google DeepMind said the release builds on earlier additions for background tasks and remote Model Context Protocol (MCP) server integration.

Gemini 3.6 Flash becomes the default

The antigravity-preview-05-2026 managed agent now runs Gemini 3.6 Flash by default, Google said. The company’s Antigravity Agent documentation also lists Gemini 3.6 Flash as the default model and describes options for model configuration.

For developers, a default-model change can reduce setup work for new deployments, while configurable model selection preserves a way to adapt an agent configuration for different tasks. Google’s documentation presents model configuration alongside background execution, remote MCP support and token-budget settings.

Hooks can inspect or stop tool use

The most consequential addition is environment hooks. Google says these hooks can be used to block, lint or audit tool calls within the managed sandbox.

Google AI for Developers’ hooks documentation says a hook can run a custom script or make an HTTP request immediately before or after a managed agent executes a tool. This gives developers points to apply their own controls around actions such as code execution, package installation or retrieval.

Pre-execution hooks could be used to evaluate a proposed action before it runs, while post-execution hooks can support logging or auditing after a tool call. Google does not describe hooks as replacing application-level security measures; rather, they offer an additional programmable control point in the managed environment.

Cost and scheduling options

Google also announced token-budget controls, scheduled triggers and free-tier access for Managed Agents. Token budgets are intended to give developers a way to constrain usage during agent runs, while scheduled triggers enable work to be initiated at set times instead of only through direct requests.

Together, the additions reflect a focus on operational control for agents that can perform multi-step tasks in a remote environment. Developers can now combine a default Flash model with configurable model settings, budget limits, scheduling and before-or-after tool-call hooks.

Google’s examples use the @google/genai TypeScript and JavaScript SDK, while the company directs Python and cURL users to its Antigravity Agent documentation.