Jihoon Jeong asks the question: if the model starts fresh each cycle, what actually compounds? He uses software developer Geoffrey Huntley’s Ralph Wiggum loop to illustrate:
If the loop’s power came from looping — from persistence of effort, from the agent grinding away at the problem — then the longer you could keep one agent going, the better it should get. The opposite is true, and every practitioner knows it. A long agentic session curdles as its window fills with dead ends and stale state; the agent gets worse with continuity, not better. The winning configuration, rediscovered by everyone who runs loops at any scale, is maximum discontinuity: kill the agent every iteration, resurrect it blank, and let it inherit nothing except what the last iteration wrote to disk. Huntley’s design wasn’t naive. It was surgical. Discard the mind, keep the files.
Fresh context is only useful if the loop can tell progress from activity. Jeong puts that burden on the verifier:
Second, one of the five decisions is load-bearing in a way the others aren’t. The verifier is the wall the whole structure hangs on. A loop repeats whatever its verifier accepts; if the verifier is strong — tests, compilers, benchmarks, anything with teeth — the loop compounds progress, and if the verifier is weak, the loop compounds output. Every experienced loop practitioner converges on the same rule: the loop is exactly as good as its stopping test. A loop with a weak verifier isn’t an autonomous engineer. It’s an expensive random walk with excellent posture.
Jeong’s answer is that the model starts each cycle from scratch, while plans, tests, commits, and code preserve progress for the next one:
The loop works, and the skeptics are right about why its working is strange. It adds no intelligence. It makes no model smarter. It rents the same brilliance every cycle at full price, extracts what it can, and throws the brilliant thing away — keeping only the residue on disk, because the residue is the only part that compounds. It works better than it has any right to, exactly as I said last time. And its characteristic failure mode is now visible at scale too, and it is not a crash. It’s a flatline. The loop keeps turning, the tokens keep burning, and the density of correct answers stays wherever the verifier pinned it — because nothing inside the system learns from one cycle to the next. The agent that finishes iteration forty is precisely as capable as the one that started iteration one. Only the pile of files has grown.
