Anthropic's Project Swap is an experiment about what happens when AI agents negotiate with other agents on behalf of people. Employees brought books they wanted to exchange and Claude-powered agents participated in a miniature trading environment.
Why the experiment is useful
The task goes beyond answering a question. Each agent had to represent a person's preferences, evaluate offers, negotiate and complete a trade.
What the experiment can teach
Agent-to-agent interaction adds a new layer of uncertainty. A user's stated preference may not fully describe what they would accept, and an agent can make a locally reasonable trade that does not match the person's broader preference.
Why human preferences are hard to encode
People often use context that is not written down. A book may have sentimental value, a preferred genre may depend on the author, or a user may care more about fairness than getting the highest-value exchange.
Lessons for product teams
- Give agents clear preference information.
- Allow users to set boundaries before negotiation.
- Require approval when a decision is hard to reverse.
- Log offers and decisions so users can review the process.
- Test edge cases where preferences conflict.