Grok Bot 六代理任务流一周拦下 37 个假完成

Ackerman @Yarilo7brigada原推发布于 2026.10.27

在 X 上打开原推

this Grok Bot terminal claims it caught 37 fake “done”s from its own bots in a week wild result. but the six-agent setup is what’s worth looking at instead of taking a bot’s word for it, the author split the process: → a dispatcher assigns tasks and attaches context from the brain → a worker does the job → an evidence agent collects command output and files → a verifier only reads and gives a verdict → a handoff agent writes the note for the next bot → a logging agent records decisions and the reasons behind them the terminal puts the whole process on one screen a task board next to the activity feed accept / reject / retry / escalate verdicts evidence checks before a task closes which brain files the bot read before starting stats on accepted and rejected tasks the author reports: tasks this week: 286 rejected by the verifier: 37 accepted on first try: 81.4% events logged: 1,964 these are self-reported numbers, not an independently verified audit and 81.4% accepted on first try doesn’t settle anything. the verifier might just be too soft what makes the architecture interesting is the ability to trace a task backwards who assigned it what context was passed what evidence was collected why the verdict was what it was who picked it up next that’s far more useful than a green task board with no history behind it if you’re building something like this, make verification transparent before the “done” rate gets impressive save this for splitting the worker from the checker, not for the number of closed tasks

这条是引用着另一条写的：x.com/yarilo7brigada/status/2113938858874470436