LLM agents in a double auction converge slower than humans, or not at all.

Replicating seminal economic experiments with LLM agents substituted for human subjects, markets populated by agents showed slower convergence toward equilibrium or none, giving less efficient allocations than human-populated markets, with substantial heterogeneity across model families and market roles. A lexical analysis of chain-of-thought traces found the decision to execute a trade rather than keep adjusting price coincides with a shift from strategic language toward urgency. The framing as an alignment dimension is right: compatibility with mechanisms designed for humans is a property nobody is testing.