Claude Managed Agents とは何か — エージェントを「長時間・非同期に動かす」ための土台

The bottleneck for capable agents is no longer the model's intelligence but the infrastructure around it. An agent needs more than a prompt: credentials to reach private systems, an identity for who it is, and a sandbox to run in. Managed agents is a set of API endpoints that handles all of this for you, offering primitives you can pick and combine, so you build your product instead of the plumbing. You open a session, your users send events, and Claude streams its work back through an event stream. It can even spawn other agents to share the work. Above all, the value is observability — a console where you watch what each agent is doing and why, which, together with sandboxing and scoped permissions, helps make them practical and governable in production.