Claude Duet 无人值守代理工程指南：七阶段流程

le0xet @le0xet原推发布于 2026.10.25

在 X 上打开原推

Anthropic put out Claude Duet Harness Engineering today. 12 minutes, official, and it finally spells out the three walls around an unattended agent: isolated browser, credential proxy, independent checks. This is f**king gold, so I rewrote it as a guide for people who ship. Seven stages, one direction: brief -> context -> tools -> observations -> verification -> repair -> delivery. The rules that make the stages hold: -> Every step declares its inputs and its required outputs. -> Tool results, artifact paths and errors get written down, always. -> Checkpoints survive a crash, so the agent resumes, not restarts. -> Output is verified against acceptance checks, not against vibes. -> A failed check goes back for one targeted repair. -> Missing evidence is labeled UNVERIFIED and shipped as such. -> Delivery means files, setup steps and the check results in one packet. On top: five runtime decisions and a worked example, CSV -> searchable comparison board -> checks -> delivery packet. Save this, then read my article on harness engineering for my AI employee ↓

这条是引用着另一条写的：x.com/slash1sol/status/2098013767627837827