Four Layers of AI Engineering

AI engineering can be mapped into four layers, each wrapping the one before it:

1. Prompt engineering. Compose a single input from role, instructions, examples, format. This got commoditized first.

2. Context engineering. The prompt is assembled from multiple sources: query, documents, memory, tool outputs, prior turns. A curator selects and compresses to fit a finite window. This is where the real skill is shifting to.

3. Harness engineering. The LLM runs inside a harness: gather context, call the model, dispatch to tools or sub-agents, verify, retry if needed. This is where most production AI systems actually live. The engineering is in the harness logic, not in the model call.

4. Loop engineering. The harness itself runs inside a larger loop that decides whether to run it again. Goal definition, budget constraints, progress checks, completion verification. The difference between “run the agent once” and “run the agent until the task is actually done.”