Decoupling the Brain from the Hands: Claude Managed Agents (Part 1 of 4) — Anthropic

This post details the architectural redesign of Claude Managed Agents, moving from a monolithic container to a decoupled system. The original architecture combined session, harness, and sandbox in a single 'pet' container, leading to debugging challenges and model-specific assumptions encoded in harnesses. The new architecture separates these into an append-only session log, a 'brain' (harness + Claude), and 'hands' (sandboxes + tools). This decoupling allows for replaceable components, injects credentials at an egress proxy (keeping them out of the sandbox), and reduces time-to-first-token. The core APIs and flow for a simple joke-teller agent are introduced, highlighting the benefits of this modular approach for agent development and maintenance.