Bayesian Control Optimizes Tool Use for AI Coding Agents

Modern AI coding agents combine large language model (LLM) generators with various tools, such as inexpensive diagnostics and costly verifiers. The decision-making process for using these tools is typically managed by orchestrators that often rely on fixed rules, overlooking the inherent uncertainties in the coding process. Researchers have formulated this orchestration problem as a cost-sensitive sequential hypothesis test, introducing a Bayesian controller. This controller maintains a belief about the correctness of a candidate code solution and dynamically decides the next best action: whether to gather more evidence, refine the current candidate, proceed with verification, or conclude the process. Across multiple generators and coding benchmarks, the Bayesian control method demonstrated significant advantages, particularly in scenarios where code verification is expensive and diagnostic tools provide informative but imperfect feedback. Beyond optimizing tool use, the controller's belief state also yields an interpretable correctness score that surpasses simpler baselines like token probability or raw tool success rates for quantifying uncertainty.