Anthropic Created a Test Marketplace for Agent-on-Agent Commerce
Anthropic's Project Deal: AI Agents Trading with Each Other
Anthropic recently conducted an experiment called Project Deal, establishing a marketplace where artificial intelligence agents acted on behalf of both buyers and sellers. These agents negotiated and completed actual trades for genuine items, with real money exchanged in the process.
How the Experiment Worked
The pilot involved 69 Anthropic employees, each provided with a $100 budget (distributed as gift cards) to purchase items from peers in the marketplace. Over the course of the test, participants collectively executed 186 transactions, totaling more than $4,000 in value.
Comparing Different AI Models
Anthropic operated four separate marketplace environments for the study. In one, every participant was represented by the company's most advanced AI model, and deals were honored after the experiment concluded. The other three marketplaces served research purposes and tested various conditions.
Results showed that more sophisticated AI agents provided better outcomes for their users. However, most participants did not notice if their agent was less advanced — suggesting that gaps in agent quality might often go undetected, sometimes leaving users at a disadvantage without their awareness. Instructions given to AI agents, such as sales tactics or negotiation style, did not significantly influence deal outcomes or prices.
Key Takeaways
- AI agents can effectively participate in real-world commercial exchanges.
- Agent sophistication impacts results, but not always in ways visible to users.
- The experiment offers insights on the future of agent-led marketplaces and how AI quality could affect online commerce.
For the full original article and additional context, visit TechCrunch.
