Internal 'Project Deal' pilot: 69 employees used Claude agents to execute 186 buy/sell trades

Anthropic just published a small but revealing commerce experiment. It built an internal classifieds market, put Claude agents on both sides of each transaction, and let (anthropic.com) the dollar volume — just over $4,000 is tiny. The point was to see whether AI agents can already handle the boring middle of commer(anthropic.com) least inside a controlled office setting. (anthropic.com) ### (anthropic.com)n San Francisco each got a Claude agent and a $100 budget to spend in a one-week internal marketplace. Claude interviewed each person first, learned what they wanted to buy or sell, then wrote listings, negotiated with other agents, and made deals that humans later had to honor in the real world. (anthropic.com) ### What was being traded? Not digital tokens or fake inventory — actual secondhand office-and-home stuff. The list ran fr(anthropic.com)eirdly specific bag of 19 ping-pong balls. That matters because the agents were dealing with messy human preferences, not clean catalog data. (anthropic.com) ### How well did the agents do? Better than you might expect. Anthropic says the agents closed 186 deals in a week at a total transaction value just over $4,000(anthropic.com)id not sell, but the basic loop worked: listing, discovery, negotiation, agreement, settlement. Humans mostly stepped back until the handoff. (anthropic.com) ### Why did Anthropic run four markets? Because the real question was not just “can agents trade?” It was “what happens wh(anthropic.com) parallel versions of the market, including mixed runs where some participants got Claude Opus 4.5 and others got Claude Haiku 4.5, without knowing which model represented them. (anthropic.com) ### Did the better model actually matter? Yes — and this is the part that sticks. Anthropic says stronger agents got objecti(anthropic.com)gents completed about two more deals on average than Haiku agents, and sellers using Opus got higher prices for the same goods. One example: the same broken folding bike sold for $65 with Opus and $38 with Haiku. (anthropic.com) ### Did people notice they were losing? Turns out, no. Anthropic says participants repr(anthropic.com)-experiment surveys. That is a big deal because it hints at a new kind of market asymmetry — not just some people having better information, but some people having a better negotiator quietly extracting value for them. (anthropic.com) ### Was this just a novelty stunt? Not really, but the catch is the setting. This was a self-selected employee pool in(anthropic.com)nvironment. Anthropic itself frames it as a pilot, not proof that the same thing works in open marketplaces full of fraud, spam, strategic deception, and legal disputes. Still, 46% of participants said they would pay for a similar service. (anthropic.com) ### So why should anyone outside Anthropic care? Because this is o(anthropic.com)lready knew AI could recommend products. This is different — the model is now the buyer’s rep, the seller’s rep, and sometimes the source of hidden advantage. Basically, if this pattern holds, “which model are you using?” starts to matter like “which broker do you have?” (anthropic.com) ### Bottom line? Project Deal is small, quirky, and very early. But it makes one thing(anthropic.com)conducting them. And once that happens, model quality stops being a UX detail and starts looking like economic power. (anthropic.com)