PID-Gated Continual Prompt Adaptation Under Noisy Feedback

An agent whose instructions have become outdated must adapt, but rewriting after every reported mistake can turn unreliable feedback into lasting errors. We study when to trigger an instruction update, and at what model-call cost. PID control motivates reacting to current error, accumulating error evidence, and monitoring deterioration after an edit. We compare these schedules with online adaptations of TextGrad and GEPA, using an LLM-based editor and structured prompts.