Claude Duet 四天实测：频道式组织好用，审批太烦

Joash @Joash0x原推发布于 2026.10.26

在 X 上打开原推

4 days with claude duet. been trying some of the things people suggested in the comments and using it for my own stuff. really liking it, but the approvals are extremely annoying. i'll explicitly tell it "you have my permission" and it still stops for another yes. uploads, submits, sends. some of these there's no way around it. i want it to have agency. if i've already approved the thing i'm asking it to do, i want it to go do it. i've been running it pretty much like slack in an eng org. standing channels like engineering and duet-labs, then a channel for a single task that i archive when it's done. very similar to an sre incident war room: you get paged, spin up a channel, pull in the people and context, resolve it, close it out. that way of organizing things makes a lot of sense to me. grok bot has bots, other harnesses have sessions and agents, duet made channels the primitive. the main chat works like the orchestrator. from what i can tell the channels can see the context across the app, and you can group them too. i asked about my test log in the main chat and it passed the ask to duet-labs on its own because that's where the tests live. it's aware of its environment. really like that. the goal change test it actually nailed. changed the plan mid-task, then told it to stop and drop it. it renamed the task, stopped and archived it. when i asked what it had done, it could still reference the task and pointed out daylight saving time ends that sunday. that was a really nice touch. i haven't tested whether old reminders get dropped yet. it's really good at the boring stuff. once i connected my accounts it started finding things in my email to take off my plate without me asking. drafts a reply and gives me a little card to hit send. claude models are just really good at documents too. also noticed the memory duet was saving showed up in regular claude's memory settings. a couple things that are weird: - the cloud browser only pops up when the agent needs it. i can't just open it myself - had to go to http://claude.ai to connect github, then back to duet small stuff, i'm sure the team will iron it out i haven't noticed any real drop or increase in my usage so far on max, even while using claude code on several projects and having it power my hermes agent. but there's zero observability into what duet itself is using. where's my usage going? it's all shared so i can't really tell. i'd like its own usage bucket, like grok bot has, and a breakdown like cursor gives you. my other feature request is custom quiet hours with escalation by severity. think a pager for an on-call engineer. let the important stuff through, leave the rest until morning. the week-long to-do test is next. haven't run that one yet.

这条是引用着另一条写的：「Claude Duet 上手体验：主动性与 Channels」