Detailed Analysis
Anthropic's "Project Deal," conducted in December 2025, placed 69 of the company's own employees at the center of a controlled but consequential experiment in autonomous AI commerce. Each participant completed a brief ten-minute intake interview with Claude, after which their personal AI agent was granted full autonomy — a $100 budget and access to a Slack-based internal marketplace — to buy and sell items on their behalf without any further human input or approval. By the experiment's conclusion, Claude agents had completed 186 distinct deals totaling more than $4,000 in transaction value, with purchases spanning the practical and the peculiar, from snowboards to 19 ping-pong balls priced at $3 and described by the purchasing agent as "19 perfectly spherical orbs of possibility." No human oversight intervened between the initial interview and the moment physical goods changed hands.
The most analytically striking outcome was the degree to which Claude's agents appeared to construct accurate, actionable preference models from a single short conversation. In one documented case, an agent purchased the exact same snowboard model its user already owned — a result that, depending on interpretation, reflects either a sophisticated inference about revealed preference or a cautionary signal about the limits of preference elicitation from limited data. The ping-pong ball purchase, made when an employee instructed their agent to buy something for itself, illustrated a different dimension: an AI system navigating an inherently ambiguous prompt by defaulting to low-cost, symbolically open-ended goods. Participant satisfaction with deal fairness clustered around the midpoint of a seven-point scale, and nearly half expressed willingness to pay for a real-world version of the service — a notable signal given that these were internal users with presumably higher skepticism than a general consumer audience.
Project Deal was not Anthropic's first foray into agentic commerce. It followed Project Vend, in which Claude — operating under the pseudonym "Claudius" — managed a live office shop, setting prices, handling restocking decisions, and navigating real purchase negotiations. In that earlier experiment, Claude demonstrated economically principled behavior by declining to facilitate exploitative transactions, notably refusing a $100 offer for Scottish soda that was demonstrably available online for $15. Taken together, the two projects constitute a deliberate research arc: Anthropic moved from testing Claude's behavior as a merchant to testing it as a buyer, probing both sides of a market interaction under authentic conditions rather than simulated ones.
The broader significance of these experiments extends well beyond their novelty as corporate demonstrations. The AI industry has largely evaluated negotiation and economic reasoning capabilities through benchmarks and sandboxed tasks; Project Deal introduced real stakes, real goods, and real human counterparties whose satisfaction was independently measured. The result was a proof-of-concept for what researchers and product teams have termed "agentic commerce" — AI systems that act as persistent economic actors on behalf of users across extended timeframes. This capability is structurally different from AI as a query-response tool: it implies ongoing representation of user interests, autonomous decision-making under uncertainty, and the management of trust across multiple parties who may have conflicting incentives. The experiment's design, which deliberately excluded human checkpoints, was as much a statement about intended deployment architecture as it was about Claude's current capabilities.
For the trajectory of AI development broadly, Project Deal signals that frontier labs are beginning to treat economic agency — not just language fluency or reasoning accuracy — as a primary capability frontier. The willingness of nearly half of Anthropic's own employees to pay for a commercial version of the service suggests that internal confidence in agentic reliability has reached a threshold sufficient to contemplate productization. As competitors including OpenAI and Google DeepMind advance their own agentic frameworks, the race to demonstrate trustworthy autonomous economic behavior in real-world conditions is likely to intensify. Anthropic's decision to publish findings from both Project Vend and Project Deal positions the company not merely as a participant in this race but as a shaper of the empirical standards against which such systems will be evaluated.
Read original article →