43. Agentic Loops - The Loop Engineering Debate

Go to my ⁠⁠⁠⁠YouTube⁠⁠⁠⁠ to see the video version of this podcast. In this episode, I explore six different things that I don't think we were expecting or ready for, but they're coming due to AI. Follow ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠@TechBackControl ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠for the infographics and visuals which accompany this episode. Agentic Loops and Loop Engineering: Who Controls the Loop?This episode explains why “agentic loops” suddenly became a hot topic in early June, sparked by Boris Cherny saying he no longer prompts Claude much because “the loops” do it, Peter Steinberger urging people to stop prompting coding agents and instead design loops, and Adi Osmani naming it “loop engineering,” which drew both “you invented the for loop” ridicule and warnings it’s only for a tiny minority. It traces the evolution from ReAct (reason–act–observe) and reflection, through Jeffrey Huntley’s “Ralph Wiggum” bash loop and Andrej Karpathy’s Autoresearch optimization loop, to today’s formal loop specs in tools like Codex and Claude Code (e.g., /loop or /goal) that define verification and stopping criteria. The key debate is control, cost, and failure modes: infinite-scale loops can burn money if mis-specified, shifting humans from “in the loop” to “on the loop” to set goals, guardrails, budgets, and success measures.00:00 Why Everything Is Loopy00:32 The Loop Engineering Spark01:50 Who Controls The Loop02:14 ReAct And Early Agents02:49 Ralph Wiggum Goes Viral03:45 Autoresearch And Optimization04:29 Modern Loop Specs In Tools05:34 Stacked Loops And Scale06:31 Infinite Failure Modes07:12 Humans In Loop To On Loop08:27 Shift Left And Control Surface09:27 What Happens Next10:15 Closing Thoughts