Already 5 months old when the scout caught it.
Managed agents
What it is
A managed agent is a cloud-hosted agent runtime a model provider operates on your behalf: the agent loop, session/conversation state, and a sandboxed execution environment (code execution, tool calls, file system) all run server-side under the provider's control, rather than being self-hosted and orchestrated by your own application code. You call an API to create, configure, and govern the agent's lifecycle; the provider handles isolation, credentials, and (in Anthropic's version) optionally lets you keep tool execution on your own infrastructure while the agent loop itself stays hosted.
Why it's emerging now
As agentic workloads move from prototypes to production, teams kept re-solving the same infrastructure problems — sandboxing untrusted code execution, persisting session state across long-running tasks, scoping credentials and network access — themselves. Anthropic launched Claude Managed Agents in April 2026 to take that off developers' hands; Google followed in May with Managed Agents in the Gemini API built on its Antigravity harness. Two labs converging on the same name for the same shape of infrastructure in the same two months is what makes this a term of art rather than one vendor's product label.
How it relates to what you know
It's the productized, server-hosted counterpart to the agent harness — the same orchestration, context management, and tool-execution concerns an agent harness handles, but run and billed as infrastructure rather than code you deploy yourself. It sits next to MCP and computer use as another piece of agent plumbing being standardized, and it's the natural next step after agent sprawl made "who is running and securing all these agents" an operational question labs now compete on answering.