Harness Engineering and Agentic Loops
How to Explain an AI Agent to Your Mum
When I was at university studying Communications, I spent a good proportion of a term trying to understand Semiotics, which is the study of how things stand for other things. My reading list ran the length of my arm in a library that smelled of old books. I love that smell, which probably dates me, since it belongs to the days when books in a library were still a thing (I imagine it is not much of a thing now… but maybe it still is?)
I have been thinking about that library a lot lately, because two phrases from my own work give me the same feeling, harness engineering and agentic loops. People say them carefully, as though you would need a reading list to follow along, which you do not.
Harness engineering might sound like a proper discipline that comes with a certification and an annual conference with a lanyard and a tote bag, but it describes the whole setup you build around a model so it can do work. That setup is the brief that tells it who it is, what it is for and where to stop, together with the tools you give it, the drawers you let it open, the memory it keeps as it works, and the information you feed it each time it runs. All of that put together is what people are calling the harness.
We could have kept it simple and called it the setup, but harness sounds cooler to clients and investors, so here we are.
When I set up an agent, I spend a lot of time mapping out the tools and the tasks I need it to do, and once its environment is in place, I make sure its job description is as clear as I can get it, because models don’t have “common sense” and the clearer you are, the better.
What about Agentic Loops? It’s not a separate thing, and it is really just one of the things the harness does. The agent has a go at something, looks at how it turned out, adjusts, and goes again until the work is good enough.
My Mom ran a loop like this every Sunday of my childhood, getting the “kinilaw” (Filipino ceviche) right. She would taste it, decide it needed something, add the something, and taste it again, round and round until it was good enough for Sunday dinner. She never thought to call it agentic, and she would have looked at me very strangely if I had.
Put the two together and that is an agent, a model you have set up properly that works in a loop until the job is done. What most of this means is a job brief and a second draft, and that is most of what these words mean.
I’m not taking a dig against the people who use these terms, but the jargon scares people away from AI. Even the capable, intelligent business owners get anxious when I mention agentic loops, because it doesn’t sound like it’s a room built for them. They have run good companies for twenty years, and there they are feeling behind, over a word that was built to sound harder than it is.
The whole reason I do this work is that the technology could bring more people in, but jargon does the opposite.
You already understand both of these ideas. You set up a harness every time you brief a new person on your team and tell them what they are responsible for and what to leave well alone, and then set them up with a working space, and you run a loop every time you draft something, read it back, and fix the parts that are wrong.
The words are new, but the thinking behind them is something you have been doing your whole working life.
All the Zest 🍋
Cien
*Cien Solon is a founder and AI transformation strategist working at the intersection of people, platforms, and power. Through [LaunchLemonade](http://launchlemonade.app), she helps organisations design AI systems that are dependable, governable, and human-centred.*
Sources and further reading
Ferdinand de Saussure, *Course in General Linguistics* (1916). The book that started the whole field.
Roland Barthes, *Mythologies* (1957). Where the theory gets pointed at wrestling matches, soap powder and steak and chips.
Daniel Chandler, *Semiotics for Beginners*. A free and refreshingly readable primer, if you ever fancy wandering in.
*: The field has founders, and they were not brief about it. Ferdinand de Saussure split every sign into a signifier, the word or sound, and a signified, the idea it points to, and handed down the founding principle that “the linguistic sign is arbitrary.” There is nothing tree-like about the word tree. All of which is a long-winded way of saying that things stand for other things.