Loops, harnesses, factories
Three words get used interchangeably and mean three different things. Getting them apart is the whole prerequisite for every decision that follows.
Ask three engineers what an “agent” is and you get three answers, all of them partly right. The word has been stretched across a chat window, a shell script, and a company’s entire delivery pipeline. That vagueness is expensive: almost every argument about whether AI development “works” turns out to be two people talking about different layers.
There are three, and they stack.
A loop
A loop is one agent doing a single job on repeat: gather context, take an action, check the result, go again until some condition is metprimary. That is the whole definition. It is the smallest unit of agentic work, and everything above it is loops stacked on loops.
The shift from prompting to looping is smaller than it sounds and matters more than it sounds. Prompting is you, in a chat window, deciding what happens next each turn. Loop engineering is you stepping out of that position and building the small system that decides what happens next — so the agent’s next turn is driven by the check, not by your attention. Huntley’s phrase for the posture is exact: sit on the loop, not in itprimary.
A loop with no stopping condition does not stop. That failure is the most common one in practice, and the reason the next layer exists.
A harness
A harness is the walls around a loop: the sandbox it runs in, the tools it can reach, the memory that survives between runs, and the gates that decide what “done” meansprimary. The loop is the behaviour; the harness is the environment the behaviour runs inside.
The formulation that made this a discipline rather than a folk practice is Böckeler’s: Agent = Model + Harnessprimary. It is worth sitting with, because of what it implies about ownership. The model is rented. It changes underneath you on someone else’s release schedule, and your leverage over it is roughly the leverage you have over any vendor. The harness is yours. It is the part you design, version, test and improve — and, over a year, the part that explains why two teams using the same model get different results.
A raw model with no harness will spin forever, cheerfully. It has no way to know it is failing.
primary There is now a preprint arguing the software-engineering gap is a harness problem — context, tools, verification, rollback — rather than a model-capability problem. Whether or not that holds in general, it is the working assumption of every team in the field report.
A factory
A factory is many harnessed loops running at once, fed by a queue of work and drained through a review gate into production, with humans owning the whole thing from aboveprimary.
The important negative: a bigger agent is the wrong mental model for a factory. Think of an org chart made of loops. The questions it raises are organisational — who feeds the queue, who owns a loop, what happens when two loops touch the same file, who is accountable when the gate passes something bad — and they do not have technical answers.
The closed loop, and the one box that resists
Draw the factory as a cycle and something useful falls out. Intent, from leadership and engineers, and signals, from incidents and users and monitoring, feed a queue. The harness picks an item and builds a change. Automated checks — CI, tests, static analysis, scanning — run at scale for negligible cost. Deploy ships it. Monitoring turns production back into signalsprimary.
Now price each box. Generation: near zero and falling. Checks: near zero, and they parallelise. Deploy: solved, a decade ago. Monitoring: solved. Every box in that diagram is close to free.
Except one. The review gate is judgment, and judgment has not got cheaper. It is the only box whose unit cost is still measured in a person’s time, and chapter 5 is about what follows from that.
What actually moved
The paradigm shift is not that machines write code. It is that the unit of work moved up a level — to the loop, the harness, and the flow between them, rather than the individual diffprimary.
That relocation is the source of most of the disorientation in the field right now. An engineer whose craft is expressed in diffs finds the diffs have become cheap and abundant, which feels like devaluation. The craft did not disappear; it moved one layer up, into designing the thing that produces the diffs and deciding which of them anyone should look at. Whether that is a promotion or a loss depends on the day.
Before arguing about whether to adopt something, name the layer it lives on. A tool that improves a loop, a discipline that improves a harness, and a reorganisation that builds a factory are three different decisions with three different costs — and only the third one changes who is accountable for what ships.