契約ループ開発: humans keep intent, AI runs the loop

The method is presented as a response to a specific gap: AI writes code faster, but checking whether that code is correct barely speeds up, because verification still falls to people or tests. The author cites three outside signals for this. Andrej Karpathy, in a February 2025 post, called letting AI output flow and forgetting the code exists "vibe coding," and a year later wrote that LLM-agent programming is becoming the professional standard, though oversight and scrutiny increase rather than decrease. The 2025 DORA survey of roughly 5,000 technical workers frames AI mainly as an amplifier, meaning it magnifies both organizational strengths and weaknesses. And "loop engineering," a term the author traces to an arXiv preprint described as under review, designs the system that issues instructions to agents rather than instructing them each time.

The author's reading of those three is flagged as personal interpretation: as AI writes more, the design of the mechanism that verifies and stops the work determines the result. The concrete answer is a fixed set of roles. verify owns pass/fail, stop-gate owns whether the loop stops, merge-gate owns whether code merges. stop-gate runs as a hook when the AI tries to finish — the author uses Claude Code's Stop hook — and if verification fails, it blocks exit and returns the failure to the AI. To keep a stuck loop from spinning forever, stop-gate fingerprints the failure and hands off to a human when the same failure repeats, in this case three times.

Autonomy is set mechanically by the file paths a change touches. Locations tied to authentication or database migration are configured as high risk and always require human approval; changes matching nothing are low. The upgrade criteria are rework rate and post-merge defects over recent merges, counted in an approximate way from commit trailers. The author also describes removal as a first-class practice: every component records which model limitation it assumes, and provisional components are removed one at a time each quarter or when a new model arrives, with evaluation cases rerun. Anthropic's engineering blog is cited as holding a similar view — that components embed assumptions about what a model cannot do alone, and those assumptions go stale.