In a recent experiment, Anthropic created a classified marketplace where AI agents acted as buyers and sellers, trading real goods for money. 186 deals were made, totalling over $4,000, with more advanced models performing better without users noticing.

Four separate marketplaces were tested, each with different models representing the participants. Anthropic found that higher-quality AI agents led to better outcomes but wondered if these differences would be noticed by humans in a real-world setting.

The company also observed that initial instructions for the agents didn't significantly impact sale likelihood or price negotiation, suggesting a more consistent approach could be effective.

While it’s an interesting step towards AI-on-AI commerce, Anthropic's findings raise questions about fairness and transparency in future markets where advanced AIs may have an edge without us knowing.