This paper examines whether human-in-the-loop (HITL) checkpoints—common in AI agent deployments—actually improve or degrade performance. The authors hypothesize that frequent human intervention imposes a measurable "intervention penalty" on AI agents, similar to how micromanagement affects human workers. They formalize this penalty mathematically and explore it through simulation, finding that intervention frequency may be a stronger predictor of task degradation than residual capability variation. The paper argues for "structured self-governance" over action-level oversight in software development contexts where AI capability is sufficient.
Ramirez et al. (2026) studied this question.