🧠Anthropic Splits Claude Agents Into Brain, Hands, and Session Log

Your agent containers can finally die without killing the agent

Anthropic rebuilt Claude Managed Agents so the parts that run your agent can fail independently. The original design crammed everything into one container. That made debugging a nightmare and recovery worse. If you're building long-running agents, this matters because the failure modes are the real product. The harness, which calls Claude and routes tool calls, now sits outside the container and treats it like any other tool. The sandbox is cattle: when it breaks, you reinitialize it from a standard recipe instead of babysitting it. The session is an append-only log of every event, living outside Claude's context window. Its getEvents() interface lets the brain pull positional slices of history instead of stuffing everything into the prompt. Credentials are never reachable from the sandbox running Claude's generated code, enforced either by bundling auth with a resource or holding it in an external vault. Watch how fast the interfaces stabilize. Anthropic says each piece can be swapped without disturbing the others, which is a promise worth testing against your own infrastructure.