Google has officially unveiled Managed Agents (preview) at I/O 2026, offering developers direct access to its underlying Agent infrastructure via the Gemini API. With a single API call, developers can instantly spin up an isolated, ephemeral Linux sandbox in the cloud — a fully managed environment running the Antigravity agent, powered by the new mid-tier Gemini 3.5 Flash model.
This new APProach eliminates complex orchestration code and environment setup. Instead of writing glue code, developers simply write two Markdown-based, veRSIoned configuration files — AGENTS.md and SKILL.md — to DeFine system roles and custom Skills, then register them as a managed agent.
Inside the cloud Linux sandbox, agents have a high-boundary ACTivity space. They can perform multi-step reasoning, task planning, call tools, execute code, read/write files, and fetch real-time data via external web access. The sandbox state can be automatically saved or restored between API calls, allowing seamless session recovery with all files and runtime states intact.
This managed infrastructure approach closely mirrors what anthropic introduced with claude Managed Agents in April 2026. Both major LLM providers are converging on a similar vision: evolving from pure API model providers into contAIner-orchestration platforms that SECurely host agent runtimes, sandboxes, and session states.
The Managed Agents preview is now available through the Gemini API. Developers can get started quickly using custom templates in Google AI Studio Playground. google also introduces a new Gemini API Agent Skill specifically for Coding Agents to build and manage their own custom agents. For enterprises, the Gemini Enterprise Agent Platform is launching in private preview.
Comments & Questions (0)
No comments yet
Be the first to comment!