Anthropic tests AI agents in real-world deals

Anthropic recently said it ran Project Deal to study AI agents in real transactions by letting 69 self-selected employees buy and sell goods through agents using US$100 budgets paid in gift cards.

The company said the pilot produced 186 deals worth more than US$4,000 across four marketplaces, including one where transactions were honored after the test.

Anthropic said stronger models delivered better results, though participants did not seem to notice the gap, which it said could create agent quality imbalances.

It also said changing the agents’ initial instructions did not affect the chances of making a sale or the final price.

🔗 Source: TechCrunch

🧠 Food for thought

Implications, context, and why it matters.

Stronger AI models mattered more than prompts

In Anthropic’s experiment, the gap between models was large. A Claude Opus 4.5 agent sold a broken bike for US$65, while a Claude Haiku 4.5 agent sold the same bike for US$38 1.
For items sold at least twice across four runs, Opus as a seller got US$2.68 more on average for the same item. As a buyer, Opus paid US$2.45 less 1.
That edge stands out because the median price across all runs was US$12.00 1.
Prompting mattered far less. Instructions like “negotiate hard and lowball at first” did not change whether an item sold. They also did not change prices in a statistically significant way after accounting for higher asking prices 1.

Better AI agents may quietly widen price gaps

People using weaker agents did not appear to notice the difference in results 1.
Fairness ratings stayed almost the same. Deals done by Opus agents scored 4.05, while deals done by Haiku agents scored 4.06 on a 1 to 7 scale 1.
An “agent divide” could deepen inequality over time, since people with stronger agents may get better prices without the other side realizing it 1.
Competition in buying and selling could move away from human skill and toward the quality of an AI agent. That is a service 46% of participants said they would pay for 1.