Abstract
With the increasing application of control theory in LLM agents, this paper introduces a novel perspective on the controlled variable. Prior works focus on controlling tool selection, inter-agent message routing, or the raw action stream. Instead, we treat context assembly itself as the controlled variable, which includes selecting prompt templates, determining the number of few-shot demonstrations, how much context to retrieve, and how many planning/verification passes to perform. This is achieved through the formal decomposition of the outer context policy $\\pi_\\phi$ and the inner frozen policy $\\pi_\\theta$.
Key Contributions
- Stability Argument: We provide a stability argument for the online controller, showing that the expected reward is non-decreasing under bounded policy change, as per Zhang et al. (2026).
- Uncertainty Calibration Analysis: An analysis of the controller's confidence against realized task outcomes ensures the reliability of its outputs.
- Empirical Evidence: The same controller is instantiated across three domains and two model providers, with released datasets, trajectory logs, and deployment recipes.
Conclusion
This paper contributes a new control-theoretic framework for context assembly, emphasizing the importance of online learning and policy control, laying a foundation for future research on LLM agents.
Blogger's Review: This paper presents an innovative perspective by treating context assembly as a controlled variable, breaking away from traditional approaches focused on tool selection and action stream control. The application of control theory offers robust theoretical support for strategy optimization in LLM agents, warranting further exploration and discussion.