This article introduces a Bayesian controller for orchestrating modern coding agents, addressing the limitations of fixed-rule systems that ignore uncertainty during tool use.
- The approach formulates orchestration as cost-sensitive sequential hypothesis testing to dynamically decide between gathering evidence, refining candidates, or verifying them.
- Evaluations across six generators and nine coding benchmarks show the method is most valuable when verification is costly and critics are informative but imperfect.
- The resulting belief state provides an interpretable correctness score that outperforms token-probability and raw tool-success baselines for uncertainty quantification.