TechCrunch: Anthropic tested a marketplace where AI agents negotiate on your behalf

Anthropic ran an internal experiment called Project Deal, in which AI agents negotiated real commercial transactions on behalf of participants. Reported by TechCrunch on April 25, 2026, the experiment involved 69 Anthropic employees using Claude models to negotiate exchanges — book trades, service swaps, and similar deals — within a purpose-built marketplace. The agents completed 186 transactions with a combined value of over $4,000.

The most significant finding was not that agents could complete negotiations, but what happened when agents of different capability levels participated in the same marketplace. Participants represented by more advanced Claude models consistently achieved better outcomes. The critical detail: participants failed to notice. Anthropic described this as a potential “agent quality gap” — a situation where one party receives systematically worse representation without any awareness that the gap exists.

Why it matters for product managers

The pilot is deliberately narrow. All 69 participants were Anthropic employees, the stakes were low, and the marketplace was designed for the experiment rather than drawn from real commercial conditions. The findings are suggestive rather than conclusive.

That said, the asymmetry finding has direct relevance for any product that deploys AI agents on behalf of users. If agent quality creates invisible outcome differences, platform designers face both a product design question and a transparency question: what should users know about the capability level of the agent acting in their interest, and how should that be communicated?

For marketplace-style products where multiple parties’ agents interact — procurement tools, contract platforms, scheduling services — agent quality becomes a variable that affects outcomes without being visible to users. Whether platforms choose to disclose this, surface agent quality as a product feature, or treat it as a background implementation detail is a design decision that will shape user trust. Anthropic’s warning about “agent quality gaps” suggests that early-stage agent products should think through this before deployment rather than after.