Loop Engineering

Addy Osmani, at Google, put a name on a practice already running: designing the loop that prompts a coding agent rather than prompting it by hand. The Lushbinary guide dates the shift to June 2026 and places «Loop Engineering» as the third layer of a stack (prompt, then context, then loop), each layer containing the previous one. Geoffrey Huntley's technique Ralph, from early 2026, was the pre-terminological proof: a `while` loop, the same prompt every pass, fresh context at each iteration, durable state on disk in `PLAN.md` and `STATUS.md`. Peter Steinberger's line, "you should be designing loops that prompt your agents", and Boris Cherny's work on «Claude Code» sit underneath the term's virality.

Four notes approach it from different seats. Shubham Saboo moves it out of engineering into product management, with five parts: trigger, action, evidence, memory, stop condition. The last is the one he calls most critical, because systems fail for lack of a clean exit. His named problem is artifact drift: the `CLAUDE.md` lengthens, the checklist swells, eval criteria change without a trace, and a month later the agent "seems worse" while the model has not regressed. Andrew Ng separates three nested loops by timescale, minutes for agentic coding, tens of minutes to hours for developer feedback, days for external feedback. Dorra Bartaguiz sums software factories as loop engineering plus harness engineering plus human supervision.

The tension nobody closes: Lushbinary holds that "done" is a claim rather than proof, verification stays human, and comprehension debt accelerates with faster loops.