AI를 다루는 기술은 세 층으로 쌓입니다 — 프롬프트(어떻게 묻는가), 컨텍스트(무엇을 아는 상태로 시키는가), 그리고 루프(누가 다음 턴을 여는가).
2026년에 더해지는 층은 루프 엔지니어링입니다 — 에이전트에게 프롬프트를 치는 게 아니라, 에이전트를 프롬프트하는 루프를 설계하는 일입니다.
Working with AI stacks in three layers — prompt (how you ask), context (what the agent knows when you ask), and loop (who opens the next turn). 2026's layer is loop engineering: you no longer prompt the agent, you design the loop that prompts the agent.
ai-toolingcontext-engineeringpillar-belief
AI를 다루는 기술은 세 층으로 쌓입니다 — 프롬프트(어떻게 묻는가), 컨텍스트(무엇을 아는 상태로 시키는가), 그리고 루프(누가 다음 턴을 여는가).
2026년에 더해지는 층은 루프 엔지니어링입니다 — 에이전트에게 프롬프트를 치는 게 아니라, 에이전트를 프롬프트하는 루프를 설계하는 일입니다.
프롬프트 엔지니어링(2023), 컨텍스트 엔지니어링(2025) 위에 한 층이 더 올라온 것입니다. Steinberger가 "에이전트를 프롬프트하는 루프를 설계하라"고 말한 것을 Alex가 학습의 세 번째 층으로 묶었습니다. 컨텍스트가 루프의 전제 — 맥락이 충분해야 루프가 스스로 돌 수 있습니다.Also in EnglishWorking with AI stacks in three layers — prompt (how you ask), context (what the agent knows when you ask), and loop (who opens the next turn). 2026's layer is loop engineering: you no longer prompt the agent, you design the loop that prompts the agent.
It sits on top of prompt engineering (2023) and context engineering (2025). Alex folds Steinberger's "design the loop that prompts your agents" into the third layer of the learning stack. Context is the prerequisite for the loop — only a well-fed agent can run the loop unattended.