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

Building robust agentic infrastructure requires moving away from monolithic, stateful containers toward a decoupled, modular architecture. By separating the session log, the agentic harness, and the execution environment, developers can treat infrastructure components as interchangeable "cattle" rather than fragile "pets." This design allows for flexible context management and improved security by routing external tool calls through a proxy, ensuring the model never directly accesses sensitive credentials. Furthermore, on-demand sandbox provisioning significantly reduces latency, achieving a 90% improvement in time-to-first-token. This approach ensures that if a component fails, the system can seamlessly recover by replaying the append-only event log, maintaining task continuity without manual intervention. These architectural principles enable developers to build more resilient, scalable, and secure agents using Anthropic’s infrastructure.