Ascertain whether agentic review induces cognitive load on a limited resource

Ascertain whether the additional iterations, retrieved context, and generated tokens observed when LLM agents review unhealthy code impose load on a limited agent resource analogous to human cognitive load.

Background

The paper observes that lower-CodeHealth code is associated with more agent turns, more retrieved context, and more generated tokens, while the number of reported findings does not increase. Although this behavioral pattern resembles the additional effort associated with difficult human code comprehension, the study does not measure any internal resource or state corresponding to cognitive load.

The authors explicitly avoid claiming that LLM agents comprehend code in the human sense and leave unresolved whether the observed expenditure reflects pressure on a bounded resource. Resolving this question would require measurements beyond token, iteration, and context-use statistics.

References

We nonetheless decline to reuse the label cognitive load for agents. Shang et al. (2026) make that move, mapping cognitive load theory’s bounded working memory onto an LLM’s context budget by direct analogy, and the analogy commits to a claim our instrument cannot support: that the difficulty arises from an exhausted resource standing between an unprocessed representation and a stable schema. Nothing we measure speaks to whether such a resource exists or is exhausted. For most PRs, our pipeline uses only a fraction of the LLM’s context limit. We therefore report the expenditure itself and leave open whether it placed load on any limited resource of the agent, since that is not something we can observe.

— CodeHealth Predicts the Token Cost of Agentic Code Review: A Pre-Registered Study of 465 Agent-Authored Pull Requests  (2610.21914 - Heander et al., 15 Oct 2026) in Section 5.3, Cognitive Load and Agentic Inference Cost