The original title is "Prompt Engineering, Context Engineering, Loop Engineering: What Actually Changed"

Short summary

The article traces how AI engineering evolved from prompt engineering (tuning one message) to context engineering (curating the full context window) to loop engineering (engineering the generate-check-retry loop itself). The key shift is that the bottleneck moved from the model to the verifier — the mechanism deciding whether an iteration is good enough to stop. The author distinguishes deterministic checks from model-graded checks, warning that confusing the two causes agent loops to stop confidently on the wrong iteration.

• •AI engineering evolved through three units of work: single prompt → full context surface → generate-check-retry loop
• •The real bottleneck is now the verifier, not the model — but what the verifier checks is the actual engineering problem
• •Deterministic checks (tests pass, no banned imports) and model-graded checks (is this good?) answer different questions; confusing them causes premature loop termination

Generated with AI, which can make mistakes.