What's this about?
Anthropic asked 201 employees to bring books, describe their reading preferences and let Claude agents negotiate swaps. The trading mechanism worked better than the agents' initial understanding of what each person wanted.
What exactly is market efficiency here?
Each received book is scored by its place in the participant's own ranking. A perfect 1.0 means everyone gets a first choice; the feasible optimum was 0.89, while stronger Opus-only reruns reached 0.88 when judged on Claude's inferred rankings.
Why does this matter?
primary research release covering 205 market runs; no broad independent discussion signal was captured.
- Claude's inferred ordering matched human pairwise preferences 61% of the time, versus 50% for random guessing.
- Model choice moved outcomes more than “ruthless” or “prosocial” instructions.
- Participants reported 7.2/10 average satisfaction, but employees are unusually AI-friendly and only Claude-family agents took part.