Anthropic 工程部落格：Managed Agents 如何把「大腦」與「雙手」解耦
Scaling Managed Agents: Decoupling the brain from the hands
摘要分析
Anthropic 2026-04-08 工程文解釋 Managed Agents 架構：把 agent 拆成 session（只能追加的事件日誌）、harness（呼叫 Claude 與路由工具的迴圈）、sandbox（執行容器）三個可獨立替換的部件。容器掛了可重建、harness 當機可從日誌復原；p50 首 token 延遲降約 60%、p95 降逾 90%。憑證放 vault 而非 sandbox，prompt injection 拿不到 token。並提醒 harness 的每個設計都隱含模型做不到某事的假設，模型升級後要重新檢驗。
- 三層拆分：session（事件日誌）／harness（迴圈）／sandbox（執行環境）
- session 當作外部上下文儲存，需要時取片段而非不可逆壓縮
- 憑證放 vault／proxy，sandbox 內的程式碼永遠碰不到 token
- 模型升級後要重新檢驗 harness 的每個假設
- 解耦後 p50 延遲降 60%、p95 降逾 90%
In this 2026-04-08 engineering post, Anthropic explains the architecture behind Managed Agents, its hosted service for long-running agent work. The agent is split into three virtualized parts: a session (an append-only event log), a harness (the loop that calls Claude and routes tool calls) and a sandbox (where code runs and files are edited), each swappable without disturbing the others. Treating containers and harnesses as cattle, not pets, means a failed container can be reprovisioned and a crashed harness restarted from the session log without losing state; p50 time-to-first-token fell about 60% and p95 over 90%. Credentials live in vaults or proxies outside the sandbox, so prompt injection cannot reach tokens. The post also argues that every harness component encodes an assumption about model limits that must be re-examined as models improve: context resets needed for Sonnet 4.5 were unnecessary with Opus 4.5.
- Three layers: session, harness, sandbox
- Session as durable external context storage
- Credentials in vaults/proxies, never in the sandbox
- Re-examine harness assumptions as models improve
- p50 latency down ~60%, p95 down >90%
關鍵字：Managed AgentsharnesssandboxAnthropicagent 架構
來源原文：https://www.anthropic.com/engineering/managed-agents
本頁為報導內容的摘要與分析，版權屬原媒體所有；請以原文為準。