We present a method that synthesizes reactive character behaviors for continuous games as compact, human‑readable programs. Current game AI practice still relies on manually authored behavior trees, state machines, and scripts, while academic reinforcement learning typically yields opaque neural controllers that are costly to train and hard to edit. To bridge this gap, we search directly over a domain‑specific language (DSL) designed for continuous‑space game policies. The language is built around reactive geometric decisions and includes higher‑order constructs such as direction maximization, which discretize the continuous behavior space into enumerable program structures. To make the search tractable, we introduce a large set of synthesis antipatterns that prune redundant program forms while preserving behavioral coverage. The search combines bottom‑up symbolic enumeration with top‑down guidance from a coding agent. Our approach, called agentic sketching, lets the agent propose a high‑level policy skeleton and then calls an enumerator to fill in local program slots. We evaluate the method on a benchmark of 14 continuous games, ranging from classic control tasks to multi‑agent football. Experiments show that pure enumeration is often more efficient than using a coding agent alone, and the combined method substantially outperforms both. The results suggest that programmatic policy search can serve as a practical authoring tool for game AI: designers specify reward functions, and the system discovers editable, effective, and often surprising behaviors.
Review: This work demonstrates the promise of integrating formal program synthesis with learning, offering an interpretable and easily iterated solution for game AI development.