Position: Agentic AI Orchestration Should Be Bayes-Consistent (Accepted ICML 2026)

arXiv·medium signal

Argues that the control layer orchestrating LLMs and tools is where Bayesian principles yield the highest return — maintaining beliefs over task-relevant latent quantities, updating from observed agent/human interactions, and choosing actions via utility-aware policies. Many high-value deployments rely on decisions under uncertainty (which tool to call, which expert to consult, how many resources to invest) where calibrated beliefs outperform heuristic routing. Provides concrete design patterns for Bayesian control in multi-agent systems.