I work towards generally intelligent partners for humans, building benchmarks and training methods for AI agents that interact with computers, the physical world, and with humans and other agents.

🌐 hao.computer · 🎓 Google Scholar · 🦋 Bluesky · 𝕏 @_Hao_Zhu

Sotopia — Open-ended social learning environment for language agents. ICLR 2024 Spotlight.

CooperBench — Why frontier coding agents lose ~50% of their solo performance when forced to collaborate. COLM 2026

AutoLibra — Turning open-ended human feedback into reliable agent metrics. ICLR 2026.

WebArena — Realistic web environment for autonomous agents. I authored browser_env, the environment layer adopted by BrowserGym, AgentLab, VisualWebArena, and HuggingFace OpenEnv. Used in production at OpenAI Operator, Meta, Google DeepMind, ServiceNow. ICLR 2024.

aact — Lightweight Python actor-model library powering asynchronous multi-agent systems.