Loop Engineering (loop-engineering-on-product) is a reference implementation of a product loop that never finishes: a fresh agent runs each round through current-state audit → research → value gate → PRD → develop → build gate → validate → ship, then hands control back to the top.
What makes it run unattended
- Two independent gates — a value critic before anything is built, and a validator after; both are separate agents that return one parseable result line
- A deterministic driver (a shell script) that only reads result lines, with bounded retry edges so every round ends in finite time
- All state in files committed to git — positioning, state, research briefs, PRPs — so each round starts with a clean context
- Trajectory monitor and deep audit every N rounds; positioning is the only slow node that can change the objective, and it needs either a human or two high-level agents in agreement
- Runs on an isolated loop branch; a human merges to live
Ships with real overnight run data — 20 rounds on a web product and 112 iterations on an iOS app — and documents in Traditional Chinese with prompts and code in English. Copy the .claude/, scripts/ and product/ folders, write one adapter, and run.