Google Gemini API Managed Agents: Antigravity, Sandboxes and Security Controls

Managed agents move agent infrastructure into the API

Google's Gemini API documentation describes a managed-agent system that provisions a hosted Linux sandbox from a single API call. Inside that environment an agent can reason, execute code, manage files and browse the web. The current documentation lists the general-purpose Antigravity agent and Deep Research as managed-agent options, with Antigravity powered by Gemini 3.6 Flash by default and configurable with supported Gemini models.

Why this matters for developers

The service reduces the amount of infrastructure a team has to build before testing autonomous workflows. Google documents Ubuntu-based environments with Python 3.12 and Node.js 22, persistent agent state with automatic VM spin-down after inactivity, and a limit of up to 1,000 managed agents. The docs also point developers to integrations and frameworks including LangChain/LangGraph, LlamaIndex, CrewAI, Vercel AI SDK, Google ADK and the Antigravity SDK.

Security remains an application responsibility

Managed agents are in public preview, and Google's guidance is explicit that teams should review agent actions and outputs before relying on them in sensitive workflows. Sandboxes are isolated at the operating-system level, but outbound network access is unrestricted by default unless developers configure an allowlist. Google also recommends trusted external tools, least-privilege credentials, short-lived tokens and verification of generated code, data transformations and configuration changes before deployment.

For teams evaluating autonomous coding, research or operations agents, the practical takeaway is that managed execution is becoming a first-class API capability, but permission design, network restrictions and human review remain essential parts of production deployment.