What happened
- On September 24, Anthropic published Project Swap, an experiment in which 201 of its employees, across six offices, brought books to trade. Each person had a Claude agent that negotiated on their behalf in an agent-to-agent marketplace.
- After a brief conversation with its owner, each agent ranked books according to their preferences with 61% pairwise agreement, versus 50% for chance and 55% for a conventional recommender system.
- 85% of the market’s efficiency loss was explained by incomplete information about what each person wanted. Negotiation between agents explained the remaining 15%.
- Agents using the Opus model reached 0.88 efficiency, versus 0.75 with Haiku. Instructing the agents to be ruthless or cooperative changed the result by 0.02. In mixed markets, Opus agents outperformed Haiku agents by 0.14.
Why it matters
- If purchases go through agents, whoever pays for the more expensive model negotiates better. According to this experiment, the gap isn’t fixed with better instructions. That turns model quality into an income advantage for the consumer who can afford it.
- For retailers and brands in Chile, the useful figure is the 85%. Value is lost in what the agent doesn’t know about its owner, not in the haggling. Whoever controls that initial conversation about preferences will control which products reach the table.
- The experiment didn’t include hostile agents. They were all Claude and they all behaved well. The real market won’t have that condition.
The number
0.02 versus 0.14. That’s how much changing the instruction moved the result, compared with how much changing the model did.
Context
The infrastructure for agents to shop is already being built. Six banks asked for buyers to be notified every time an agent pays, and software started working as a buyer in processes that used to go through people.
What’s next
- Anthropic acknowledges the study’s limits: participants who trust Claude more than average, no financial incentives and fixed market rules.
- No timelines announced for a version with agents from other providers or with adversarial agents.
Bottom line
It had already been measured that two AI engines name the same brand in 81% and 43% of their answers. If, on top of that, the agent doing the buying depends on the plan its owner paid for, the storefront stops being the same for everyone.
Sources
Edited by Rodrigo Cornejo. How we select and verify: who writes these notes.
