Scaling Managed Agents: Decoupling the brain from the hands

Source: Anthropic Engineering Date: 2026-04-08 URL: https://www.anthropic.com/engineering/managed-agents

Summary

Anthropic’s Managed Agents service moved from a monolithic “pet” container architecture to a decoupled design separating the brain (Claude + harness), hands (sandboxes), and session (append-only event log). Stateless harnesses communicate via a standardized execute(name, input) interface and recover via a wake(sessionId) mechanism. The redesign cut time-to-first-token by ~60% at p50 and over 90% at p95.

Implications

The managed agents / platform thread. This is Anthropic publishing the internal infrastructure architecture for their own agentic product — the event log + wake mechanism is the durability primitive that makes multi-session agents reliable at scale. This directly informs how any persistent-agent framework should be designed.

Stateless harness as a design principle. Moving the harness outside the sandbox and making it stateless is a specific architectural decision with broad implications: any harness that holds state in-process is a reliability liability. The OS analogy (stable interfaces outlasting implementations) is the framing Anthropic is pushing.

Security model. Credentials external to sandboxes — bundled at init for Git tokens, vault-proxied via MCP for OAuth — is a concrete security pattern for any shop running Claude agents against real services. This is the right pattern and Anthropic is now publishing it as the reference architecture.