AI Loops Can Run Forever — Too Bad That's the Problem

The article argues that loop engineering for AI agents is not about making models retry, but about designing verifiers that give each cycle a genuine direction toward a desired state. The author identifies four conditions for convergence: a target state, an observable current state, a precise way to make local changes, and an external stopping rule. A key insight is that loops are tuned to their stack — bespoke loops that work on one codebase often fail elsewhere, which explains why published loops feel magical to some and useless to others. The piece draws parallels between how humans decide work is 'done' (tests, deadlines, reviews) and how AI systems need similarly external, structured signals rather than open-ended 'make it better' prompts.