Do agent representation failures or negotiation skills cause market shortfalls?

Hitzig and colleagues at Anthropic report that in Project Swap — a 201-person barter market where Claude-powered agents traded books on participants' behalf — "the market fell short mostly because of the information agents lacked about their participants, rather than because of how they traded." Once agents reached the trading floor, "the agents traded well"; the shortfall from the best possible allocation was concentrated upstream, in how well a short intake conversation captured what each person actually wanted. An agent's ranking of ten sample books, built from "a five-minute chat," matched its participant's own ranking on only 61% of pairs — a figure the authors call "surprisingly good for such a short conversation," but still a wide gap from full knowledge of someone's preferences.

The authors frame this as a representation problem bearing on the two questions the post opens with: when is an agent fit to act for someone, and what rules keep an agent acting loyally for its person from harming everyone else. Because "most of the shortfall from the optimal outcome was due to the agents' inability to represent preferences from a short intake," they argue the fix belongs before the trading floor, not on it — in longer or better intakes, and in tests that let a person check, and correct, how well an agent understood them before trusting it to act alone. Their own prototype for this was quiet: after intake, participants ranked a handful of books, and the team compared that ranking to Claude's guess, without showing participants the score. The authors note some of the gap may be irreducible — participants described their reading taste as inherently vague ("I don't even fully know what I want") — so better elicitation narrows but may not close the shortfall.

This locates the bottleneck earlier than other papers in the library place it. Do shopping agents respond more strongly to price and rating cues? and Can a seller infer what a buyer really wants to pay? both find the buyer-side agent is the vulnerable party once trading starts — manipulable by a counterparty, or legible to one. Project Swap's agents instead fail earlier, before any counterparty is in the picture, simply at the task of hearing their own person correctly. That upstream framing also gives empirical shape to Must AI monitoring precede incentive alignment as agents evolve?: Project Swap's intake-ranking comparison is a working instance of the monitoring step that paper argues must come before an agent can be trusted to act on incentive alignment alone.

The study does not establish how much a longer or multi-turn intake would close the 61%-of-pairs gap, nor whether the result generalizes past Anthropic employees trading books — the authors flag that their participants were unusually willing to trust Claude, and unincentivized to rank carefully. It also says nothing about adversarial counterparties, since "all agents were built from Claude production models, post-trained to be polite and largely cooperative." The implication the authors draw is narrower than agents are ready to trade for people: it is that before any marketplace rule-making matters, someone has to solve, and let a person verify, whether their agent understood them at all.