The Great Loops Debate — Dex Horthy, Geoff Huntley, Ian Livingstone, Greg Pstrucha, @insecure-agents

Oxford Style Debate: There is, or is not, a delta between the hype behind loops and what actually works in practice. Team No Delta (pro the way we do loops today) The hype around loops is valid and loops work well today in practice. Loops today can be a silver bullet and result in outsize productivity gains, and marks an important step up the autonomy curve towards real software factories. Ian Livingstone Geoff Huntley Team Delta (anti the way we do loops today) There is a delta between the hype behind loops and what actually works in practice. The way we are doing loops today is wrong. Loops are not a silver bullet and there is no magic. The hype is outrunning the discipline "Stop writing loops, start writing control loops." A bare repeat-the-agent loop isn't magic. The leverage comes from the Kubernetes-style reconciliation around it: read current state → read desired state → one incremental change → repeat. Dex's tell when shown a fake loop: "where's the recur condition?" (Jun 21) A software factory can run the mechanical, spec-gated, test-covered slices unattended; it cannot autonomously decide whether it built the right thing. Dex Horthy Greg Pstrucha Main Debate Loop History - Why now as the inflection point and not some of the earlier ones? Loop Anatomy - What makes a good loop? Loop Future - Given what we’re seeing with loop usage now, are we well positioned for software factories? If we can’t use loops well today how do we expect to operate software factories? Appendix Research https://x.com/AnatoliKopadze/status/2068328135611822149?s=20 https://x.com/ericzakariasson/status/2070493377267646797?s=20 https://x.com/MilksandMatcha/status/2069838072515281386?s=20 https://x.com/AnatoliKopadze/status/2070156017262793008?s=20 https://ghuntley.com/loop/ https://ghuntley.com/ralph/ https://www.anthropic.com/institute/recursive-self-improvement Anthropic's Absorption of the Ralph Loop Verifying Agents in GitHub 0:00 Introduction and format explanation 0:43 Introduction of the debaters 4:05 Team No Delta (Ian & Jeff) opening arguments 6:53 Team Delta (Dex & Greg) opening arguments 10:28 Rebuttal and initial stance 15:32 Main Debate: Loop History and Inflection Points 16:20 Security, alignment, and goal-seeking agents 19:24 Why loops became more usable today 24:40 Context rot and context engineering 27:10 Loop Anatomy: What makes a good loop? 30:30 Preventing agent cheating and verification 32:23 Convergence engineering and loop slop 36:13 Economic viability and token spend 39:07 Multiplayer agents and shared memory access 43:21 The "just write loops" advice critique 46:36 Scaling, autonomy, and pragmatism 52:31 Closing statements and final thoughts Viral Quotes & Pull Quotes: (16:55) "As these models get better, the most important thing to remember is they actually become higher goal-seeking and higher capable in terms of finding exploits to achieve their ultimate goal." (20:25) "These LLMs generate code better than you can actually hire for. It's sad but true." (22:38) "The models are drunk, right? You can't trust them. But like, we accept that. We engineer away those failure domains." (36:31) "I don't think they fail quietly. I think they fail very loudly, especially when you're looking at your bills." (46:58) "Don't throw away all the things we've learned. Don't go out of your way to cast aside this decades-long career of software engineering that we as a community have built up."