Multi-agent system created work just to finish it
The author's GPT-6 Astra agent system saw task counts surge within a week while demanding more attention, and only settled after adding a rule that every new task must point to an external outcome.
i accidentally built an AI employee that was creating work just to finish it for a week, the numbers looked great. more completed tasks. more files. fewer blockers. but my GPT-6 Astra setup somehow required more of my attention than before. I opened the activity logs and found the reason. one agent would discover missing information and create a research task. the research agent returned a document. that document triggered a review. the review created corrections, which triggered another task. nothing was technically broken. every agent was doing exactly what its role told it to do. the system had simply learned how to stay busy. so I added a rule I should have started with: every new task had to point to an external outcome. a customer gets something. a decision becomes possible. a real blocker disappears. or an artifact I actually need gets finished. if it couldn’t point to one of those, the task died there. the activity graph immediately got smaller. so did the number of “completed” tasks. but more useful work reached me. that’s when I stopped measuring my agents by how much they do. an AI company can automate busywork so well that it accidentally manufactures more of it
This post quotes another post: x.com/lummox_eth/status/2100288151897981255