Anthropic's Managed Agents: Decoupling AI Planning from Execution for Scalable Long-Horizon Tasks

Anthropic discusses the evolution of agentic AI systems, specifically the concept of "Managed Agents" — a hosted service in the Claude Platform that runs long-horizon agents by decoupling the planning "brain" from the execution "hands." The article details how different Claude models (Sonnet 4.5 vs Opus 4.5) exhibit different behaviors like "context anxiety," requiring adaptive harness designs. The core thesis is that agent infrastructure (harnesses) must be model-agnostic and evolve independently from the models themselves, and that Managed Agents provide a stable interface layer that outlasts any particular implementation.