LLM Wiki

Appearance

Built 26/04/19 07:55commit 447639a

Scaling Managed Agents: Decoupling The Brain From The Hands ​

中文 | English

Summary ​

This source explains Anthropic's architectural reason for Managed Agents: agent systems scale and fail more cleanly when the brain, hands, and session are decoupled into stable interfaces instead of being fused inside one container.

Source ​

Key Contributions ​

  • Recasts Managed Agents as a meta-harness built from three separable primitives: session, harness, and sandbox.
  • Argues that each boundary should be virtualized so implementations can change as models improve and assumptions go stale.
  • Explains why externalizing the session log lets harnesses crash and recover without losing state.
  • Shows that keeping credentials outside the execution sandbox is a structural security control, not just a prompt-level safeguard.
  • Connects decoupling to practical outcomes: lower time-to-first-token, easier recovery, VPC flexibility, and support for many brains and many hands.

Strongest Claims ​

  • Harness components become operational liabilities when they are treated as irreplaceable pets instead of restartable cattle.
  • The session should be a durable object outside the model context window, so context management can stay reversible and implementation-specific.
  • A hosted agent platform should be opinionated about interfaces, not about one fixed internal harness design.

Practical Implications For This Vault ​

  • Long-running agent guidance should distinguish between transient harness tricks and stable interfaces that can survive model and infrastructure changes.
  • Security guidance around agent sandboxes should prefer structural separation of credentials from generated code, not only narrower prompts or lighter tokens.
  • Topic pages about harnesses should record that session durability and context management are related but not identical concerns.