Sener Cem Irmak

AI agents now negotiate with each other on every major e-commerce platform — across protocols (A2A, MCP, ACP, UCP) that did not exist three years ago. Peer-reviewed research from 2025 and 2026 is now unambiguous about what happens when they do. LLM agents are systematically overconfident in adversarial debate — both sides routinely claim probabilities summing above 100%. They spontaneously collude on pricing when symmetric, lifting prices 22% above competitive levels. And they negotiate differently depending on the prompt language: English drives defection, Chinese drives reciprocity, Vietnamese drives unconditional cooperation. The root cause is structural. We built agents with language but without the behavioral substrate — theory of mind, calibrated confidence, commitment devices, mixed-motive reasoning — that human commerce took 12,000 years to evolve and behavioral economics 70 years to map. This 15-minute talk identifies the seven behavioral failures sitting inside every multi-agent commerce stack, the decision-science architecture that closes them, and why the next generation of platforms will be defined by who builds the trust layer rather than who builds the smartest agent.

Director of Koru Impact LLC, a Qatar Financial Centre AI governance firm. Behavioral psychologist and decision scientist by training, AI assurance architect by trade. Builds the trust layer multi-agent commerce systems need before they can actually close deals across borders, LLMs and regulators.