Build, Don't Play: LLM Research Agents Create a World-Class Hearthstone AI
This presentation explores how AutoTavern, a system of six specialized LLM research agents, autonomously developed TavernAgent—a Hearthstone Battlegrounds AI that rivals the game's top human players. Over 20 days, the system constructed a full-content simulator, trained a PPO policy through 23 iterative versions, and deployed it on the live ranked ladder where it reached approximately 15,000 rating. The key finding: LLM agents excel at building specialized AI systems through structured research processes, but perform poorly when asked to play the game directly—revealing a fundamental distinction between construction and execution.Script
Can language models build what they cannot play? AutoTavern reached the top of the Hearthstone Battlegrounds ladder by constructing a specialized AI system, yet those same language models fail catastrophically when asked to make the game decisions directly.
The system coordinates six specialized research agents that write to typed, versioned memory stores rather than conversing freely. Five verification gates enforce discipline: builds must pass tests, models stay paired with their simulator version, comparisons require multiple seeds, delivery needs unseen-seed confirmation, and crucially, live games validate the simulator itself.
TavernGym implements all 978 game objects from patch 36.6.3 in a Rust engine that runs 190 to 280 complete eight-player games per second. TavernAgent is an 18 million parameter entity transformer trained with PPO through 23 delivered versions, improving from 1.9 percent first-place rate on simulator exams to 41.5 percent.
Deployed on the live ranked ladder, the agent climbed from 6,000 rating to approximately 15,000 across 392 games. The final five versions achieved 32.6 percent first places and 91.3 percent top four finishes, with an average placement of 2.37 compared to 2.34 for the server's ten highest-rated players.
TavernBench reveals the limitation: when prompted to play directly, the strongest language model reaches only 63 percent accuracy compared to TavernAgent's 91 percent. Every evaluated language model except one falls to zero percent real-time accuracy under the one second decision budget that live play demands.
The distinction is precise: language model agents are effective architects of specialized systems but poor real-time controllers. If you want to explore how structured research processes unlock capabilities that direct prompting cannot, visit EmergentMind.com to dive deeper into this work and create your own video explainers.
