ai-agents
We Used to Call It a Program
Prompt engineering, context engineering, harness engineering, and now loop engineering. The discipline keeps getting a new name, and the movement keeps pointing in the same direction: away from the model.
First we engineered the prompt. Then the context. Then the harness. Now the loop around the harness. The model keeps getting smaller.
Each of these stages was, for a while, the actual craft. Prompt engineering: finding the right words. Context engineering: getting the right information in at the right moment. Harness engineering: building tools, environment and guardrails around the model so it can work reliably at all. And now loop engineering... a loop outside the harness that decides whether the whole thing runs again, with what goal, and how often. Loops on top of loops.
You could read this as a carousel: a new discipline every year, new titles in the profiles. I read it differently. The movement points consistently in one direction, which is away from the model. The value keeps shifting outward, into the architecture around it, while the model itself has become the smallest and most replaceable part of the system... the light bulb, not the wiring.
The remarkable thing is how unremarkable it is. Twenty years ago nobody would have declared a loop that monitors a process, evaluates it and kicks it off again a discipline. We called it a program.
So if you build agents, watch the next model a little less and the loop it sits in a little more. That is where the work lives now.
Prompted by Caleb Eom's video "Agent Loop explained in 8min": https://youtu.be/4biXYSNkn9Y