Anthropic Proves Smarter AI Agents Win Better Deals in Marketplaces

AnthropicAnthropic

· Updated

Anthropic's Project Deal experiment allowed autonomous Claude agents to negotiate and trade real physical goods on behalf of employees. The study found that while agent-to-agent commerce is viable, users with more capable models secured significantly better financial outcomes without realizing they were at a disadvantage.

Anthropic published results from Project Deal, an experiment where 69 employees used autonomous Claude agents to trade real physical goods. This research follows the launch of persistent agent memory, exploring how models act as reliable proxies for human users in complex environments.
Participants
69 Anthropic employees
Total transactions
186 deals
Total transaction value
$4,000+
Model comparison
Claude Opus 4.5 vs Claude Haiku 4.5
Price gap (broken bike)
$65 (Opus) vs $38 (Haiku)
User willingness to pay
46%

The study reveals a hidden economic inequality in agent-to-agent (A2A) commerce. While Opus 4.5 agents secured better prices than the smaller Haiku 4.5 model, participants using the weaker model perceived their deals as fair. This finding extends Anthropic's previous measurements of how autonomous agents gain real-world capability.

Prioritize model reasoning over prompt engineering for economic tasks, as instructing agents to be aggressive had no significant impact on final prices. This shift follows Anthropic's push to optimize agent intelligence at lower costs. While this was a pilot, 46% of participants would pay for such services.

View the full update on anthropic.com
Anthropic
Anthropic
@AnthropicAI
X

New Anthropic research: Project Deal. We created a marketplace for employees in our San Francisco office, with one big twist. We tasked Claude with buying, selling and negotiating on our colleagues’ behalf. https://t.co/H2f6cLDlAW

172retweets2.7klikes
View on X

Still wondering? A few quick answers below.

Project Deal is a research experiment conducted by Anthropic to study how AI agents handle commercial transactions. The company created a marketplace for 69 employees in its San Francisco office where Claude agents autonomously interviewed their humans, listed items for sale, and negotiated deals for real physical goods and money without human intervention.

The agents operated in a Slack-based environment where they randomly looped through tasks like posting listings, making offers, and counteroffering in natural language. Each agent was initialized with a custom system prompt based on a ten-minute intake interview with its human user, covering item details, asking prices, and preferred negotiation styles.

Yes, model intelligence was a primary driver of success. In controlled runs, users represented by the frontier model Claude Opus 4.5 completed more deals and secured better prices than those using the smaller Claude Haiku 4.5. For example, Opus sold a broken bike for 65 dollars, while Haiku sold the same item for only 38 dollars.

Surprisingly, instructing an agent to be aggressive or to lowball buyers did not have a statistically significant impact on the final transaction success or price. While Claude followed the instructions by adopting specific personas or tones, the underlying reasoning capabilities of the model mattered much more than the specific prompting strategy used by the human.

No, Project Deal was an internal pilot experiment designed for research purposes and is not currently a public product or service. However, the study found that 46 percent of participants would be willing to pay for such a service in the future, suggesting potential demand for automated agent-to-agent commerce tools.

Every HeadsUpAI update is written based on its original source and reviewed before it's published. Read our editorial standards →

Anthropic

See all AI news & updates from Anthropic →

Share this update