Grok Bot 共享大脑架构把周成本从 412 降到 38 美元

Ackerman @Yarilo7brigada原推发布于 2026.10.26

在 X 上打开原推

this Grok Bot terminal claims it cut a bot team’s spend from $412 to $38 a week wild result. but the shared brain setup is what’s worth looking at instead of letting every bot start from zero in its own chat, the author split the process: → the chief takes the task and reads brain.md first → a lookup agent checks whether it’s been done before → a specialist does only what isn’t already in the brain → a verifier demands proof before “done” → a librarian files the result into the right folders → a logging agent writes to log.md who did what and why the terminal puts the whole process on one screen the brain’s folder tree next to the activity feed read / reuse / run / skip decisions duplicate checks before any task runs token spend per task, before and after a counter of results pulled from the brain instead of redone the author reports: starting spend: $412 a week ending spend: $38 a week reused results: 61% log.md entries: 2,317 these are self-reported numbers, not an independently verified calculation and low spend doesn’t settle anything. a cheap team might just be doing less what makes the architecture interesting is the ability to trace a result backwards where the task came from which brain files were read why it wasn’t redone who checked the result where it was written that’s far more useful than a pretty savings number with no history behind it if you’re building something like this, make the memory transparent before the savings screenshot gets impressive save this for the memory split, not the weekly savings

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