Reader change: Instead of repeatedly typing “What now?” the reader learns to design a loop that specifies an activation trigger, an observable numeric stop condition, an independent evaluator, and a default failure action. With that design the AI runs autonomously each day; the article supplies a concrete, ready‑to‑use blueprint (tested in Japanese and English) that the reader can copy and deploy.
Evidence supporting the claim:
- Boris Cherny (Anthropic) stating he no longer writes prompts but writes loops that prompt the agent (Source 1).
- Addy Osmani’s definition of “Loop engineering” as replacing the human prompt writer with a system (Source 1).
- suwash’s three‑layer model separating prompt writing, context design, and loop design (Source 1).
- delba_oliveira’s taxonomy of start/stop modes (goal‑driven, time‑driven, event‑driven, etc.) (Source 1).
- Anthropic’s 800‑hour, 97 % self‑supervision experiment showing large gains when loops are correctly designed (Source 1).
- Replit’s public account that keeps hypothesis selection, implementation direction, evaluation criteria, and release decisions in human hands (Source 1).
Remaining uncertainty / limitation: The loop approach only pays off for repetitive, well‑defined judgments; it is not cost‑effective for one‑off tasks. Its success still hinges on the human‑crafted stop criteria and evaluator logic—if the evaluator crashes or is mis‑configured, the system must have an explicit default‑failure policy, otherwise low‑quality outputs could be emitted. Moreover, the Anthropic 97 % result is tied to a specific experimental setup and has not been independently reproduced, so the figure cannot be universally generalized.
답글 남기기