Context Assembly as the Controlled Variable: A Control-Theoretic View of Harness Policies for Frozen LLM Agents

This paper frames context assembly as the controlled variable for frozen LLM agents. Instead of modifying model weights or directly controlling tool actions, an outer policy selects prompt templates, few-shot demonstrations, retrieval volume, and the number of planning or verification passes around an inner frozen policy. It formalizes this decomposition, discusses a stability condition based on bounded policy changes and non-decreasing expected reward, and analyzes calibration between the controller’s confidence and realized task outcomes. The supplied abstract emphasizes formal framing and evidence rather than introducing control theory to agents generally.