Kyle Mistele, CEO of HumanLayer, argues that the dominant "Ralph" pattern of AI coding loops—a single bash loop piping a prompt to generate massive PRs—is dangerously misaligned with real-world engineering, as blind loops produce "40,000-line PRs" that no human can review, work only for solo developers or non-critical systems, and are prohibitively expensive in token cost and later debt of bad code. Citing Matt PCO's observation that each line of bad code costs far more in the agent era, Mistele advocates replacing blind loops with control-theoretic loops: incremental, measurable, human-steerable mechanisms. He notes that even inside frontier labs, blind loops cause real problems, citing that "it took six months to fix a terminal flicker in Claude Code" while a smaller team rebuilt a renderer faster. Mistele imports ideas from control theory, which has four components: Sensor, to measure the current state; Controller, to decide the next action; Actuator, to apply the change; and Reference, the target state. He walks through a production loop his team runs at HumanLayer to migrate an RPC API to the Effect library, culminating in a downloadable skill template and a call for engineers to treat loops as carefully tuned systems.