When two large language models transact as commercial agents, do the cognitive biases each model inherits from its training data persist, wash out, or amplify? We answer with seven controlled experiments — bilateral negotiation, decoy effect, information overload, private and common-value auctions, strategic exploitation, prompt-based debiasing, and continuous double auctions — across three frontier models and 8,280 runs.

We find selective and consistent vulnerability: agents play near-equilibrium in private-value auctions and resist information overload, but anchor more strongly than humans in bilateral bargaining, are systematically exploitable by an informed counterparty (mean surplus extraction of $6.47), and collapse allocative efficiency from ~98% to under 20% when biased participants populate a multi-agent market. Generic debiasing prompts fail; specific named-bias warnings provide partial mitigation.