Papamarkou et al: agent orchestration should be Bayesian

"The control layer of an agentic AI system (that orchestrates LLMs and tools) is a clear case where Bayesian principles should shine," a group of thirty authors led by Theodore Papamarkou write in a position paper at ICML. Their target is not the language model itself but the code around it that decides which tool to call, which expert to consult, or how many resources to invest.

The split is deliberate. Making the LLM itself into an explicitly Bayesian belief-updating engine remains "computationally intensive and conceptually nontrivial as a general modeling target," the authors write. The orchestrator, by contrast, is where beliefs over task-relevant latent quantities can be maintained, updated from agentic and human-AI interactions, and turned into actions. "Coherent decision-making requires Bayesian principles at the orchestration level of the agentic system, not necessarily the LLM agent parameters," the paper argues.

The abstract stops at design patterns and concrete examples. It does not publish benchmarks against non-Bayesian orchestrators. Two researchers in our tracker had already shared the link.