MerchantBench: the best LLM reached 27.3% of human net assets over a simulated year.

arXiv 2607.28956, topping HuggingFace Daily Papers today with 74 upvotes, grounds 365 simulated days of e-commerce in 98,843 real product records, forcing agents to coordinate sourcing, pricing, cash management and order handling under mixed-latency feedback. Humans finished around 217,610 RMB; GPT-5.6 Sol reached 40,890 under ReAct and 52,930 under Hermes. The failure taxonomy is the part to steal: "operational coherence" decay, where activity just trails off, and "strategic coherence" breakdown, where goals drift and the agent stops updating on evidence. Humans sustained 100% engagement. LLMs ranged 10.6–66.1%. If you're shipping anything long-running, instrument engagement rate as a first-class metric.