Anthropic ran a controlled market experiment with 201 employees using Claude-powered agents. Agents traded books on a decentralized floor after short chats about preferences. Claude’s book rankings matched human rankings on 61% of pairs. The market worked well but was limited by incomplete agent information. Stronger models produced more efficient markets than different instructions. Most participants liked their final book. Participants said they would trust Claude with a third of their yearly book budget. The experiment builds on Project Deal, a prior agent marketplace test. Centralized and decentralized market designs were both tested.
What it unlocks
Agents can now negotiate trades in a decentralized market based on user preferences.
- 201 participants
- 61% pairwise agreement on book rankings
- 61% agreement vs. 50% random
- 216 median words in intake chat
Sources