shunyu-yao

Shunyu Yao — researcher at openai; the person behind two of the wiki’s load-bearing ideas: reasoning-as-action (react) and the utility problem (agent-evaluation).

Work the wiki tracks

• CALM (2019) — text-game agents on GPT-2; per his own account, “the first agent in the world built via pre-trained language models”. Millions of RL steps per game, no transfer — the empirical anomaly that motivated ReAct (2026-04-27-the-second-half-of-ai).
• ReAct (2022) — added reasoning to the agent’s action space; the primitive underneath essentially every modern agent-loop. Cataloged as early as Weng‘s 2023 survey (2026-06-04-llm-powered-autonomous-agents), which also surveys his Tree of Thoughts (2023) under task-decomposition.
• tau-bench (2024) — user-simulator-in-the-loop benchmark; his own counter-example to the autonomous-eval assumption (agent-evaluation).
• The Second Half (2025) — the thesis essay (2026-04-27-the-second-half-of-ai): the recipe (pretraining + scale + reasoning-as-action) generalizes, benchmarks saturate in months, the field’s bottleneck moves to evaluation design — the utility problem.

The through-line

Each step follows from the last: CALM’s failure to transfer → reasoning-as-action as the fix (react) → the fix works so well the recipe industrializes benchmark-solving → therefore evaluation, not method, is the open problem. The wiki’s reasoning-models and post-training pages carry the training-side consequences; agent-evaluation and long-running-agents carry the eval-side ones.

Sources

• 2026-04-27-the-second-half-of-ai — authored.
• 2026-06-04-llm-powered-autonomous-agents — his ReAct and Tree of Thoughts surveyed as core planning/acting methods.