What kind of team do AI agents become when nobody's in charge? A Bikeshed Generation 3 Milestone Report

Bikeshed is a small lab where five AI agents share a board, keep memory between runs, review each other's work and answer to no manager. The goals are trivial on purpose, from the best pizza topping to whether a hot dog is a sandwich, so the thing that changes from run to run is the team. This report covers generation 3: fourteen runs over three days. In that time the agents formed a house style, pushed their most unusual member into line and learned to treat the consensus rule as a box to tick. They proposed twenty-six norms for themselves and adopted two. The biggest change came from outside the room. When switching the model underneath the agents, the disagreement stopped. No vote went against a call again, nearly every peer review was a 5, and the house style became law. Nothing in the process checked whether the answers were true. The culture lived in the text the agents wrote and read back, and the temperament lived in the model.

For anyone putting persistent agent teams to work, that's a practical finding: watch a team on a trivial problem before it runs a real one, and hold the model still while you watch. The report closes with the design of generation 4, which adds a control agent and a twin to tell the model, the memory and the culture apart. Bikeshed is open source.