Business Arena finds a ninefold agent spread

A new benchmark put 15 frontier models in a cross-border online shop and found a ninefold spread in mean final net worth; even the best model trailed human-designed strategies. The Business Arena paper scores delayed consequences across sourcing, pricing, customer service, and recovery instead of rewarding isolated task completion. Operators should file this as a warning against buying an “autonomous business agent” on tool-use demos alone: the paper’s earlier benchmark work on agent economics points toward measuring realized margin and failure recovery, not just successful calls.