OpenAI has published research detailing 'emergent tool use' discovered by agents training in a simulated hide-and-seek environment. The experiment involved multi-agent co-adaptation, where hide agents and seek agents developed progressively complex strategies. Initially, agents simply ran and hid.

However, as training progressed, hide agents learned to use boxes to build shelters, and seek agents learned to use ramps to break into those shelters. This led to a series of six distinct strategies and counter-strategies, including hide agents locking doors and seek agents using mirrors to see around corners. Notably, some of these complex behaviors were emergent—meaning they were not explicitly programmed but developed naturally as agents adapted to each other.

This research is significant because it demonstrates how simple reinforcement learning in a multi-agent setting can produce complex, tool-using behavior without human intervention. It suggests that future AI systems could potentially discover novel solutions to problems simply by competing against other agents in a simulated environment.