Do trading algorithms follow the rules economists designed for them?

Researchers compared how LLM agents perform in different matching markets—some with centralized rules (like stable roommate algorithms), others with free negotiation. LLM agents proved far more honest about their preferences than human subjects, yet traditional game-theoretic guarantees didn't always hold: the TTC mechanism, designed to prevent lying, didn't actually elicit more truth-telling than a simpler alternative. The upshot: classical matching theory works better with AI agents than humans, but the gap between theory and practice remains.