Compiled Agency: Frontier General-Purpose Coding Agents Build Winning Game Players from Bare Interaction — from Flappy Bird to StarCraft II and Civilization
Year: 2026
Type: preprint
Abstract: LLM agents have repeatedly struggled to convert knowledge of a game into competent play, even when researchers build the agent around the model---supplying perception, memory, skill libraries, planners, or executable-policy scaffolds. Rapid progress in coding agents raises two sharper questions: can frontier models now win games at all, and---more fundamentally---can they win them unaided, building the entire player themselves? We introduce Gauntlet, a develop--freeze--evaluate framework that ports games, private and public, from small JavaScript arcades to full commercial-scale titles, behind one uniform and deliberately bare contract: a general-purpose coding agent, used exactly as shipped, receives a game description, a raw observation/action interface, and an empty policy file---no strategy, no algorithm, no architecture. In a single autonomous session the agent experiments with the live game and engineers a standalone controller; we freeze the result and score it on held-out instances with zero model calls during play. On an unpublished procedural roguelike, held-out success spans 0--86% across deployed systems and exposes a sharp generational threshold: every observed session of one newest-generation system outperforms the best session of its immediate predecessor. At full-game scale, a compiled raw-API controller defeats every fair StarCraft II built-in AI and two cheating variants, and single-session programs win complete games of Civilization (Freeciv) by total conquest on held-out seeds---at modest rates, against the built-in novice AI, yet a first: no prior language-agent system had won full games of this genre standalone, without per-turn model calls and a hand-crafted tactical layer. Frontier coding agents begin to track, though not yet master, long-horizon strategy. The frozen programs are inspectable: winners contain live simulation, persistent state, calibrated doctrine, and finishing logic, and generalize through observation rather than seed lookup; fresh-session variance makes the distribution of artifacts, not any champion, the measured object. We call this capability compiled agency: development experience compiled into a persistent executable agent whose architecture is built by the model, not the researcher.
Source: arXiv (Cornell University)
Authors: Xiao, Joey, Huang, Haonan
Institutions: New York University, Princeton University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204699366