The experiment, called Project Deal, is the latest in a series of internal commercial pilots Anthropic has run on its own staff. It follows Project Vend, where a Claude agent ran a small office vending operation. Project Deal went further, putting agents on both sides of each transaction and running a full exchange cycle that included listing, counteroffer and close, all in natural language.

How Project Deal Works

Each participant started with a brief intake interview. Claude gathered details on what they wanted to sell, minimum acceptable prices, buying interests and any instructions on negotiation style. Those responses were translated into custom system prompts that defined each AI agent’s behavior throughout the experiment. Agents were then deployed to four parallel Slack channels and given no further instruction. They posted listings, messaged counterparts, haggled over price and executed deals. No human signed off mid-run.

Anthropic ran four simultaneous versions of the experiment, two powered entirely by Claude Opus 4.5, the company’s then-frontier model, and two using a randomized mix of Opus and Claude Haiku 4.5, a smaller model. One version was designated “real,” meaning participants exchanged the physical items their agents negotiated. Anthropic didn’t reveal which run was real and which model represented whom until the experiment concluded.

The agents handled deals covering a snowboard, lab-grown rubies, a folding bicycle, dog-sitting time and a bag of 19 ping-pong balls. Post-experiment surveys showed participants broadly satisfied with what their agents bought and sold. When asked whether they’d pay for a similar service in the future, 46% said yes.

Why Model Capability Changes Deal Outcomes

The experiment’s more significant finding came from the mixed-model runs. Agents running on Opus completed roughly two more deals per participant than those on Haiku. When the same item was sold by an Opus agent in one run and a Haiku agent in another, Opus fetched $3.64 more on average. A lab-grown ruby sold for $65 under Opus and $35 under Haiku. The same broken folding bicycle went for $65 with Opus and $38 with Haiku.

The disadvantage was invisible to the people experiencing it. Participants rated deal fairness at roughly the same level regardless of which model represented them, TechCrunch reported. Satisfaction scores for Opus and Haiku users were statistically indistinguishable. Anthropic noted the uncomfortable implication directly: in real-world agent markets, participants on the losing side might not know they’re worse off.

Aggressive negotiation instructions had no statistically significant effect on sale likelihood or final price. Model quality mattered. Prompting strategy didn’t.

“We suspect we’re not far from more agent-to-agent commerce bubbling up in the real world, with real consequences,” Anthropic said in its project summary.

What It Means for Payments

For payments firms and merchants, Project Deal puts a concrete data point behind what has been a largely theoretical conversation. Agentic commerce relocates decision-making from humans to software acting under human-defined constraints. That shift doesn’t replace existing payment rails, but it does change what sits above them. Pricing, negotiation and checkout move to the agent layer.

That creates a readiness problem on the merchant side. PYMNTS Intelligence found that nearly 80% of acquirers say they’re at least somewhat prepared for agentic commerce, but a much smaller share of merchants deliver consistent payment experiences across channels, a prerequisite for any system where autonomous agents transact across environments.

Mladen Vladic, head of product for payment networks at FIS, told PYMNTS the shift is already past the experimentation phase. “This is a transformational inflection point in the industry,” he said. “Not only in this country, but globally.”