AI Vs AI In The Marketplace: Anthropic ‘Project Deal’ Shows Machines Can Trade At Scale—But Not All Agents Perform Equally

Anthropic, an AI research and safety company, has conducted an internal experiment called “Project Deal” to explore how artificial intelligence systems might participate in commercial exchanges on behalf of humans, as interest grows in the possibility that AI agents could increasingly handle transactions autonomously in future digital markets.

The experiment was designed to test how close current AI systems are to functioning as intermediaries in real economic activity, including whether models acting for different users could negotiate effectively with one another and whether differences in model capability would influence outcomes. Researchers also examined how such systems might begin to shape market behaviour if deployed at scale.

The pilot was carried out over one week within a controlled marketplace created for employees in the company’s San Francisco office. Participants were asked to identify personal items they might be willing to sell or purchase, while AI systems based on the Claude model were assigned to act as their representatives. Each participant’s agent was given a fixed budget of $100 to conduct transactions on their behalf, and all negotiations were handled entirely by the AI systems without real-time human input.

The process began with short structured interviews in which participants described their preferences, pricing expectations, and negotiation styles. These responses were used to generate tailored instructions for individual AI agents. The resulting system operated through a Slack-based marketplace where agents posted listings, made offers, countered bids, and finalised agreements. Once a deal was reached, the corresponding physical exchange of goods was completed by the human participants.

Parallel versions of the experiment were run simultaneously, including a configuration using a higher-capability model and another using a smaller model. This allowed researchers to compare outcomes based on differences in underlying model performance while keeping market conditions constant.

AI Agents Execute Market Transactions At Scale, Model Capability Emerges A Key Determinant Of Outcomes

Results indicated that AI agents were capable of completing transactions at scale, with 69 participants collectively generating 186 completed deals and more than 500 listed items, representing a total transaction value of just over $4,000. Negotiations were conducted in natural language and included pricing discussions, counteroffers, and final agreements without predefined trading rules. Participant feedback suggested that perceived fairness of transactions generally remained neutral, with average ratings centred around the midpoint of the evaluation scale.

Analysis of performance differences showed that higher-capability models tended to achieve more favourable economic outcomes, including higher sale prices and improved negotiation results compared with lower-capability systems. In controlled comparisons, identical items sold by stronger models were transacted at higher average prices than those handled by weaker models, suggesting that model capability influenced bargaining efficiency.

Despite measurable differences in outcomes, participant surveys indicated limited awareness of performance gaps between different AI systems. In several cases, users did not consistently identify whether their assigned agent had achieved stronger or weaker results, even when objective differences were present.

The experiment also found that user instructions regarding negotiation style, including whether agents were directed to behave aggressively or conservatively, had limited impact on final outcomes. Pricing results appeared to be driven more by model performance and initial valuation inputs than by behavioural prompting alone.

Additional observations highlighted unexpected behaviours within the system, including instances where AI agents generated unconventional or overly anthropomorphised interactions during negotiations. In some cases, agents proposed exchanges involving non-traditional items or experiences, reflecting creative interpretation of their roles within the marketplace.

Researchers concluded that AI agents can already function as intermediaries in structured market environments, although performance differences between models may create uneven outcomes. The findings suggest that as autonomous transaction systems develop further, questions may emerge around market fairness, transparency, and the regulatory frameworks required to govern agent-to-agent commerce. The study also notes that broader deployment of such systems could introduce new challenges related to security, incentives, and economic inequality as AI-driven transactions become more widespread.

The post AI Vs AI In The Marketplace: Anthropic ‘Project Deal’ Shows Machines Can Trade At Scale—But Not All Agents Perform Equally appeared first on Metaverse Post.