An AI can win the negotiation and still buy the wrong thing

An agent in Anthropic’s Project Deal bought its human a snowboard identical to one they already owned. The choice matched the person’s taste with uncanny precision. It also exposed a gap in the brief: a purchasing agent needs to understand what is already at home.[2]

That problem is becoming more consequential. The Financial Times reported on 18 October that Anthropic was seeking advice from law firms about how agents should negotiate, citing two people familiar with its thinking. The question includes representing a client’s interests while respecting fair outcomes.[3]

A good trade starts before the bargaining

Anthropic’s September book exchange, Project Swap, helps separate the problems. It involved 201 employees. Claude inferred their reading preferences from brief chats; among the 188 who also supplied independent rankings, it ordered pairs of books as they did 61% of the time, against 50% for chance.[1]

The researchers scored outcomes on the humans’ rankings, from zero for their last choice to one for their first. The best feasible allocation scored 0.89. An optimal allocation using Claude’s inferred preferences, still judged by human preferences, scored 0.60. Actual agent bargaining averaged 0.55 across 80 controlled reruns.[1]

Thus, about 85% of the measured shortfall came from imperfect preference information. Only the remaining share came from the trading process. Those are results from a controlled book market among Anthropic employees, using cooperative Claude models; they do not establish how agents will behave against hostile counterparties.[1]

Better bargaining still had measurable value in the earlier Project Deal. Its four parallel markets included mixtures of Opus 4.5 and Haiku 4.5. Across completed transactions, the researchers’ model estimated that Opus sellers secured $2.68 more per item and Opus buyers paid $2.45 less than their Haiku counterparts, controlling for the item and market run. These estimates include simulated runs; only one all-Opus market determined the physical exchanges. A price advantage can therefore be real without settling whether the purchase itself was wise.[2]

A signature answers a different question

The current Agent Payments Protocol specification illustrates another piece of the emerging infrastructure. It uses signed checkout and payment mandates to demonstrate an agent’s authority to complete a transaction. In its autonomous mode, a user approves constraints and an agent subsequently assembles a purchase within them. The specification explicitly leaves the process of determining what the user wants outside its scope.[4]

That division is sensible: a merchant needs evidence that a purchase was authorised. But a valid signature cannot make an incomplete shopping brief complete. A system may faithfully honour a spending ceiling while choosing something the buyer would reject if asked.[4]

Consider a hypothetical agent buying a camera lens. A price ceiling and a preferred brand leave several questions unresolved. Must it fit a camera body already owned? Is weight more important than maximum aperture? Would the buyer accept a used lens with a shorter warranty? A discount on an incompatible lens has no useful saving to offer. These are decisions to settle before granting purchasing authority.

A useful product could turn those uncertainties into a small set of choices: two or three candidate purchases with the trade-offs made explicit, plus a reason for asking. The aim would be to establish which attributes the buyer will trade away and which are absolute requirements. The same record could then inform the purchase constraints and provide a basis for checking the eventual order.

A useful shopping agent must justify the cost of briefing and checking it with a purchase the customer values. That includes knowing there is already a snowboard at home.[2]

Report an error or suggest a correction