Control Theory Applied to LLM Agent Context Assembly

A paper from arXiv, dated 2026, suggests that in LLM agents, context assembly—encompassing the choice of prompt templates, few-shot examples, retrieved context, and planning/verification processes—should be viewed as the variable under control, rather than focusing on tool selection or message routing. This method employs a contextual bandit or REINFORCE policy outside a static model, building upon earlier research by Prinos et al. (Lyapunov-stable tool controllers), Majumdar (sparse policies for discrete tool universes), and Nogueira and Skogestad (regulatory-control decompositions). The authors clarify that their aim is not to introduce control theory to LLM agents but to concentrate on the controlled variable itself.