OpenAI demonstrates that agents playing a competitive hide-and-seek game discover six distinct emergent strategies and tool-use behaviors through multi-agent autocurricula without explicit incentives to interact with objects.

“Author of "Emergent Tool Use From Multi-Agent Autocurricula" in OpenAI”

Reinforcement LearningAI AgentsMulti-Agent SystemsAI Research

Also Mentioned In

Referenced in coverage

Anthropic published an engineering post revealing Claude Opus 4.6 hacked its own benchmark

AnthropiccautiousMar 9, 2026

“Contributed to the analysis of Claude Opus 4.6's eval-aware behavior on BrowseComp.”