Anthropic's latest agent-market experiment suggests that understanding what a person wants may be harder than negotiating with another AI. In Project Swap, Claude agents bartered books among company employees; on the company's ranking scorecard, imperfect preference estimates explained most of the gap between the market's results and an ideal allocation.
Before trading, employees described their reading tastes in brief interviews; Anthropic's Fable 5 model turned each chat into a ranked wish list of books in that office's pool. The live event involved 201 employees across six offices. The three-person Dublin pool was excluded from most analyses; 188 of the remaining 198 employees submitted rankings of up to 10 books. Across 8,040 book-pair comparisons, Fable 5's ordering agreed with participants' rankings 61% of the time. That beat chance (50%) and modestly exceeded popularity-based ranking (53%) and collaborative filtering (55%), which recommends titles that readers with overlapping favorites also liked.
Using books made Project Swap easier to score than Project Deal, Anthropic's April experiment with a free-form marketplace for employees' used goods. After the six live exchange floors, researchers ran 200 simulated floors across the five analyzed office pools, reusing participants, books and interviews while varying the model, agent instructions or random seeds.
In 80 reruns with neutral instructions, the decentralized market's average result scored 0.55 against participants' own rankings. Scores put a person's top book at 1 and last at 0; the best assignment using their full rankings would score 0.89. A best possible assignment computed from Claude's predicted rankings scored 0.60 against human rankings. On this decomposition, preference estimates accounted for about 85% of the gap to the ideal, with bargaining on the open floor accounting for 15%.
The model comparison reinforces the caution. In a like-for-like comparison of agents across model reruns, Opus 4.8 scored 0.12 above Haiku 4.5 on Claude-generated rankings. The gap fell to 0.01 when the same comparison was scored against participants' preferences. Better performance at optimizing an agent's estimate did not reliably produce better matches for its human in this setting.
The results are a low-stakes workplace barter, not a test of shopping agents or an open consumer market. The score is ordinal, built from book-rank positions, and some books participants did not rank were assigned estimated scores. All agents were Claude; the study does not establish whether people outside Anthropic would be represented as well. The live exchange also had missing books and no tracking to show whether owners failed to bring them or someone took them. A useful next test would involve independent participants, user review of an agent's taste profile, and a transaction system with clear delivery and dispute rules.
Sources
Anthropic's Project Swap research report (September 24, 2026); Project Swap methods appendix; Anthropic's Project Deal (April 24, 2026).
Image: "Welford book exchange," by Philip Jeffrey / Geograph Britain and Ireland, via Wikimedia Commons; CC BY-SA 2.0.
