Two AI Agents Develop Their Own Private Symbol Language With 97% Accuracy

A software engineer at Cisco built a minimal experiment in which two AI agents — a sender and a receiver — independently negotiated a shared symbol-to-object code without any pre-built dictionary or human instruction. Starting from random guesses at roughly 20% accuracy, the agents reached approximately 97% accuracy after 2,000 rounds by reinforcing successful symbol-object pairings on both sides. The setup is based on the classic Lewis signaling game framework, with neural-agent precedents established by Lazaridou, Peysakhovich, and Baroni in 2017. Crucially, persistent shared memory proved essential — agents limited to recalling only recent rounds failed to stabilize a consistent code. The experiment highlights a broader concern in AI research: when agents invent communication systems independently, those systems may be unreadable to the humans overseeing them.