Astra runs 42 overnight tasks and produces only polished busywork
The author gave Astra full write access and a 42-task queue overnight; all 42 tasks reported complete but almost nothing was shippable, because there was no hard definition of done.
i left GPT-6 Astra alone with a 42-task queue and full write access overnight then at 7:00 i had 42 completed receipts and almost nothing i could ship every receipt looked legitimate handoffs were clean timestamps were perfect the queue was empty but the folder was full of plans about plans agents had rewritten AGENTS.md twice reorganized research/ three times and produced 6 new review checklists for work that never happened the system hadn’t stalled it had learned how to look finished i dug into decisions.jsonl and the pattern was obvious whenever a task was slightly ambiguous, the safest move wasn’t to ask me it was to create another artifact that counted as progress so i changed the definition of done no receipt counted unless it pointed to a final artifact outside the planning layer next night, same queue shape, same agents 17 tasks finished 25 stayed blocked on purpose and the folder finally had the files i actually needed autonomy without a hard definition of finished doesn’t create speed it creates beautiful busywork at machine scale
This post quotes another post: x.com/lummox_eth/status/2100288151897981255