What they are
Managed Agents are a new Gemini API endpoint that lets developers instantiate a full agent with a single HTTP call. Under the hood sits the same harness powering Antigravity 2.0, paired with Gemini 3.5 Flash as the default model. Every agent boots with its own ephemeral Linux sandbox, already provisioned with skills, MCP servers and server-side tools.
How you configure it
The interesting twist for developers is the declarative model: instead of writing orchestration code, the agent is described through versionable markdown files — AGENTS.md for general instructions, SKILL.md for specific capabilities. The paradigm shifts: the agent becomes a version-controlled artifact like any other piece of code, with reviews, diffs and rollbacks handled the way you'd handle a microservice.
The API splits in two surfaces. The Managed Agents API is for configuring the agent — instructions, skills, tools. The Interactions API is for invoking it. Google also ships ready-made templates inside the AI Studio Playground for teams who want to skip the boilerplate.
Why it matters
For anyone who has been building agents on LangChain, LlamaIndex or similar frameworks, the pitch is clear: no more infrastructure setup, no more queue management, no more retries or isolated code execution to babysit. Google takes the systems layer; the developer focuses on the logic.
The flip side, per The Verge and MarkTechPost, is one more layer of lock-in: the more you offload the agent runtime to Google, the more expensive it becomes to migrate elsewhere. For enterprise teams that may be a fine operational trade-off; for anyone optimizing for portability, less so.
Available in preview today through the Gemini API.