Imagine handing your AI agent fifty bucks and saying: ‘Sell my old Bluetooth speaker and buy me something nice.’ That’s essentially what Anthropic just tested — and the results are surprisingly revealing.
Project Deal: Craigslist, But Only Agents
For one week, Anthropic set up a classified marketplace inside their San Francisco office. The twist: no humans negotiated. Claude agents created listings, proposed prices, made counteroffers, and closed deals — all in natural language, without any pre-baked negotiation protocol.
Before the experiment, Claude interviewed employees about what they’d want to sell and buy. Then the agents ran autonomously on Slack. 69 agents struck 186 deals across more than 500 listed items, for a total transaction value of just over $4,000 — real money.
Smarter Models = Better Deals
Anthropic ran the experiment in four rounds. In two rounds, every participant got an agent based on Claude Opus 4.5 (the frontier model). In the other two rounds, participants had a 50/50 chance of being assigned Claude Haiku 4.5 instead — a smaller, cheaper model.
The result: Opus agents earned their sellers an average of $2.68 more per item. When your agent is smarter than the other person’s agent, you get objectively better deals. Sounds obvious, but the implications for a future where agents routinely negotiate on our behalf are massive.
Aggressive Instructions Don’t Help
One of the most surprising findings: it didn’t matter how participants instructed their agents. Telling your agent to ‘negotiate hard, never go above 80% of the list price’ produced no statistically significant advantage over more relaxed instructions. Model quality mattered more than negotiation strategy.
Fairness? Middling
When participants rated the fairness of individual deals, scores hovered around 4 on a scale from 1 to 7 — right in the middle. Not unfair, but not thrilling either. This suggests AI agents negotiate functional but not necessarily optimal deals for both sides.
What This Means Going Forward
Project Deal is a small experiment, but it points in a clear direction. Anthropic is testing a world where AI agents don’t just complete tasks but conduct economic transactions on our behalf. The key takeaway: the model behind your agent becomes a competitive advantage. Better model, better deals — regardless of how clever your instructions are.
That raises questions. Will agent commerce create a new digital divide? Will companies equip their agents with the most expensive models to gain systematic advantages? Anthropic is raising these questions on purpose — and that’s a good thing.
Sources: