Anthropic Study: Preference Accuracy, Not Negotiation Skills, Limits Agentic Market Efficiency

Anthropic researchers conducted Project Swap, a controlled barter economy experiment where Claude-based agents negotiated book exchanges on behalf of 201 employees. The study aimed to determine whether agent negotiation skills or the accuracy of preference representation drives market efficiency in agentic economies.

Preference Representation vs. Market Design

The experiment revealed that preference representation is the primary bottleneck. Claude agents achieved a 61% pairwise agreement rate with users' actual book rankings, outperforming random guessing (50%) and collaborative filtering (55%). However, this imperfect understanding led to significant efficiency losses:

• Utilitarian optimum: 0.89 (theoretical maximum efficiency).
• Actual decentralized market efficiency: 0.55.
• Loss attribution: 85% of the efficiency gap was due to Claude's inaccurate preference estimation, while only 15% resulted from negotiation dynamics or market design flaws.

Even when applying optimal market rules like Top Trading Cycles (TTC), efficiency remained at 0.60 when based on Claude's rankings, confirming that better market structures cannot compensate for poor preference data.

Model Performance and Agent Behavior

Model capability significantly impacted outcomes, with stronger models achieving higher efficiency scores when evaluated against Claude's own rankings:

• Haiku 4.5: 0.75
• Opus 4.8: 0.88
• Sonnet 4.5: Scored between Haiku and Opus.
• Fable 5: Close to, but lower than, Opus.

Agent instructions had a negligible effect compared to model upgrades. Agents instructed to be ruthless (self-interested) scored only 0.02 higher than prosocial agents (cooperative), whereas upgrading from Haiku to Opus improved scores by 0.12. Prosocial agents were twice as likely to accept lower-ranked books to maximize overall satisfaction.

User Trust and Observability

Participants indicated willingness to delegate about a third of their annual book budget to AI agents. Trust was heavily influenced by the accuracy of the initial intake summary; participants who perceived no omissions in the summary were more likely to delegate budget.

The study highlighted the critical need for observability. One participant reviewed the agent's full action log and discovered it had held a desired book for an hour before trading it away, prompting them to purchase it directly. Researchers concluded that in agentic markets, transparency into the negotiation process is as vital as the final outcome for maintaining user trust.