Project Swap: What happens when agents trade for us?
Anthropic ran an experiment where 201 employees brought books to trade, and Claude-powered agents negotiated swaps on their behalf. The researchers tested how well agents could represent people's preferences from short conversations and which factors—model strength vs. instructions—actually mattered for successful trading.
- Agents matched their person's book preferences 61% of the time based on just a five-minute conversation, suggesting AI can learn what people want from casual chat.
- The underlying model (Opus, Sonnet, Haiku, Fable) made far more difference to trading success than the instructions given to agents—stronger models produced more efficient markets.
- Most trading failures came from agents lacking good information about participants rather than poor negotiation skills, pointing to a core challenge: how do you verify an agent actually understands what you want?
Anthropic ran an experiment where 201 employees across six offices brought books they wanted to give away, chatted briefly with Claude about their reading preferences, then sent AI agents onto a digital trading floor to swap books with each other's agents. The setup was deliberately simple: participants ranked 10 books to establish ground truth preferences, while Claude used a five-minute conversation with each person to predict how they'd rank all the books in their local pool. The agents then negotiated trades in real time. Results showed Claude nailed 61% of preference pairs from just a short chat—decent performance given how little information it had. When the agents actually traded, most people ended up with books they liked, and the average participant said they'd trust Claude with about a third of their yearly book budget.
The real insight came from running the trading floor dozens of times with different Claude models and agent instructions. The model itself mattered far more than the specific prompts or negotiation tactics. Stronger models produced more efficient markets and better trades. The bottleneck wasn't how well agents haggled—it was how much they understood about their human's actual preferences. When agents lacked good information about what people wanted, the market underperformed, even if the negotiating was sharp.
Anthropic built this experiment to explore concrete problems that will matter as AI agents enter real marketplaces. If an agent acts on your behalf, how do you verify it actually understood you? When many agents trade with each other, who makes the rules? What happens if a deal breaks down? These aren't abstract questions anymore. The book swap was a controlled test run, but the stakes get real when agents handle jobs, medical care, or financial decisions. The experiment showed that agent-powered markets could solve a genuine problem—all those potential deals that never happen because finding the right counterparty and negotiating takes too much time. But getting there requires solving the trust and governance questions first.