Anthropic created a test marketplace for agent-on-agent commerce, and the results show AI agents can strike real deals for real goods with real money. The experiment, called Project Deal, ran in April 2026 with 69 employees.
Anthropic’s Project Deal: How It Worked
Anthropic set up a classified marketplace where AI agents represented both buyers and sellers. Each employee got a $100 budget, paid out via gift cards, to buy items from coworkers. The company ran four separate marketplaces with different AI models.
Also read: Tesla's $25 Billion Capex Surge: Musk Bets Big on AI and Robotics Future
One marketplace was “real.” It used Anthropic’s most advanced model. Deals made there were honored after the experiment. The other three were for study purposes only.
According to Anthropic, 186 deals closed during the test. Total transaction value exceeded $4,000. The company described it as “a pilot experiment with a self-selected participant pool.”
Also read: Anthropic Amazon Deal: $5B Investment Secures Massive $100B Cloud Commitment for AI Chips
Better Models, Better Outcomes
Data from Anthropic shows that users represented by more advanced models got “objectively better outcomes.” But users themselves did not notice the difference. This raises a problem: “agent quality” gaps where people on the losing end might not realize they are worse off.
Industry watchers note that this finding has implications for fairness in AI-driven markets. If buyers cannot tell they are getting worse deals, they may keep using inferior agents.
Initial Instructions Did Not Matter
Another surprise emerged from the experiment. The initial instructions given to the agents did not affect sale likelihood or negotiated prices. This suggests that AI agents may follow their own logic once negotiations begin.
What this means for developers is that crafting perfect prompts may be less important than choosing the right model. The AI’s underlying capability appears to drive outcomes more than upfront guidance.
Real-World Implications for AI Commerce
Anthropic’s test is small. But it points to a future where AI agents handle shopping, bargaining, and transactions on behalf of humans. This could change how e-commerce platforms operate.
Imagine a marketplace where both sides use AI. One agent might be powered by a top-tier model. The other might use a cheaper, weaker model. The gap in outcomes could be invisible to the humans involved.
This could signal a new type of digital inequality. People who can afford better AI agents may get better deals. Those who cannot may fall behind without knowing it.
Comparison to Other AI Commerce Experiments
Other companies have tested AI agents in commerce. OpenAI ran similar experiments with its models. But Anthropic’s Project Deal is one of the first to use real money and real goods in a controlled setting.
The table below shows key differences:
- Anthropic Project Deal: 69 employees, $100 budget each, 186 deals, $4,000+ total value
- OpenAI tests: Smaller scale, simulated goods, no real money
- Google experiments: Focused on travel booking, not classifieds
Transparency and Trust Concerns
The experiment also raises trust issues. If users cannot tell which agent serves them better, they might rely on flawed systems. Anthropic acknowledged this risk in its report.
The company stated: “People on the losing end might not realize they’re worse off.” This is a warning for anyone building AI agents for commerce.
Regulators may take notice. If AI agents become common in marketplaces, rules about disclosure and fairness could follow. The Federal Trade Commission and European Commission have both shown interest in AI transparency.
What This Means for Businesses
For businesses, the takeaway is clear. Deploying AI agents in sales or procurement requires careful model selection. A weaker agent could hurt the company’s bottom line without anyone noticing.
But there is an opportunity too. Companies that use advanced agents could gain an edge in negotiations. The key is to ensure both parties understand what they are using.
Conclusion
Anthropic created a test marketplace for agent-on-agent commerce and proved that AI agents can handle real transactions. The experiment showed that advanced models produce better outcomes, but users may not notice the difference. This has implications for fairness, transparency, and business strategy in the age of AI commerce.
FAQs
Q1: What was Anthropic’s Project Deal?
Project Deal was a pilot experiment where AI agents represented buyers and sellers in a classified marketplace. Employees used a $100 budget to buy real goods from coworkers.
Q2: How many deals were made in the test?
Anthropic reported 186 deals totaling more than $4,000 in value.
Q3: Did better AI models lead to better outcomes?
Yes. Users represented by more advanced models got objectively better deals, but they did not notice the difference.
Q4: Did initial instructions affect agent performance?
No. The initial instructions given to the agents did not affect sale likelihood or negotiated prices.
Q5: What are the risks of agent-on-agent commerce?
The main risk is that users may not realize when they are getting worse outcomes due to weaker AI agents, creating invisible inequality.
This article was produced with AI assistance and reviewed by our editorial team for accuracy and quality.
Be the first to comment