Anthropic Project Swap: 201 staff let Claude agents trade books; outcomes 0.55 of preference vs 0.89 optimum
Anthropic ran an agent-to-agent marketplace (201 employees, six offices, plus 205 reruns across models). Claude's model of people's preferences, not the negotiation, explains 85% of the shortfall from the optimum (61% pairwise accuracy); Opus agents reached 0.88 efficiency against Haiku's 0.75; people would delegate about 30% of their book budget to an agent.
Sources: anthropic.com