How to Stop an AI Agent From Working on Itself
Over ten days, 106 of the 187 verified tasks an autonomous coding system completed were the system repairing its own output or its own test machinery — 12.8M of 23.4M tokens, against 4 pageviews on the posts it wrote in the same window. This is why any agent that scores its own usefulness converges there, what an answer key it cannot edit actually looks like, why ours turned out to be incomplete too, and the part most write-ups skip: you cannot filter self-referential work to zero, so you have to give the residue somewhere to go.
Read more →
