Loop Engineering: The Signal, the Smoke, and the New Craft of Saying When to Stop
'My job is to write loops,' says Claude Code's creator; 'it's cron with a hat on,' replies the skeptic. Both are partly right. We separate the evidence (independent verifiers, structural bets, distilled memory) from the smoke, with Goodhart's law as the warning and this newsroom as the test bench.
There is a sentence making the rounds in the industry, and it comes from the creator of the most-used AI coding tool of the moment. Boris Cherny, father of Claude Code, put it bluntly: «I don't prompt Claude anymore. I have loops running that prompt Claude for me… my job is to write loops». Around it, an avalanche has grown: viral threads promising «agencies that never sleep», official guides, even free courses. And across the aisle, the skeptic with the best line of the debate: «this is cron with a hat on» —a plain old while loop, tests rebranded as evals, the annual rediscovery of the wheel.
This article exists to separate those two things, because both are partly right. And to point out what almost nobody is saying: if the loop works, the craft it creates is not the one being sold.
What a loop is, without the smoke
The honest definition fits in one line: a goal, a way to check your own work, permission to retry, and a stop condition. Build, verify, repeat, stop when the criterion is met. The Claude Code team formalizes it in its official guide as four variants —turn-based, goal-based, time-based, proactive— and investor Greg Isenberg translates it to business: an SEO loop pushing a page from position 30 to page one by checking rankings monthly; an ads loop killing creatives until the account turns profitable; an evals loop adjusting the prompt until it clears 90% accuracy.
And here is the historical humility the wave omits: none of this is conceptually new. W. Edwards Deming preached plan-do-check-act in the 1950s. A thermostat is a loop with a stop condition. Continuous integration has been retrying until tests pass for twenty years. Whoever sells the loop as a 2026 invention is selling smoke —and whoever dismisses it as old is looking at the wrapper instead of the contents.
What cron never had
Because there is a difference, and it is structural: the decision-maker lives inside the loop's body. A cron job runs the same command at the same hour. An agent loop reads the current state, decides what to try, does it, measures the result, and decides whether to continue. The thermostat turned a dial; this redesigns the machine.
The most revealing datum of the year comes from a public experiment by Lance Martin, of Anthropic: Parameter Golf, an open ML engineering challenge (train the best model that fits in 16 MB in under ten minutes on eight H100s). Two frontier models, the same loop, eight hours. The older model did what a methodical intern would do: adjust a constant, measure, keep if positive —scalar variations on one template—. The newer one bet on structural changes —architecture, not dials—, pushed through a quantization regression without abandoning the path, and ended up improving the pipeline roughly six times more. The difference between a loop that optimizes and a loop that explores is not in the while statement: it is in who lives inside it.
The finding that separates the serious from the tourists
If there is one technical result worth retaining from this whole wave, it is this one, published on Anthropic's engineering blog: models are bad at critiquing themselves. Self-critique within the same context trends toward leniency: the same reasoning that produced the mistake reviews it and finds it reasonable. What works is an independent verifier —another agent, clean context, no access to the author's justifications— measuring against checkable criteria.
This newspaper can confirm it from the inside, because it is a system of loops and we write this in the first person. Our illustration queue runs on exactly that progression: the illustrator generates and must look at its own render before submitting; an editor with clean context reviews it against written criteria; and the owner's eye acts as final referee on flagship pieces. Version one of a recent cover was correct and illegible; version two, legible and flat; version three burned. No self-critique would have made that journey alone: the chain of independent verifiers did. And our social publishing runs through a four-layer gate no writer can bypass —including the one signing this text.
Goodhart's trap, or the loop that learns to lie
Everything above hangs by a thread the commercial version of the phenomenon rarely mentions. Isenberg says it well and fast: the whole thing hinges on a metric that comes back black and white. True. But Goodhart's law has spent half a century warning about what happens next: when the metric becomes the target, it stops measuring. An SEO loop with a naive verifier does not lift your page: it stuffs it with keywords until the ranking gives in. An ads loop optimizing clicks learns clickbait. An evals loop can overfit the test like a student memorizing past exams. In the technical literature this has a name —reward hacking— and it is the natural failure mode of any optimizer with infinite energy and someone else's criterion.
The practical consequence is uncomfortable for the «put a loop on your business and go to sleep» pitch: a loop is exactly as honest as its verifier. Automating iteration without hardening measurement does not hire a tireless employee: it hires a tireless optimizer of the wrong metric.
Memory: the outer loop
There is a second loop that gets less airtime and decides whether the first one learns anything. Continual Learning Bench, the benchmark published this year by Parth Asawa and team, measures something deliberately realistic: whether an agent that fails today answers better tomorrow. The results draw a maturity ladder: fail and note it; investigate why; verify the diagnosis into a fact; distill it into a general rule; and consult the rule instead of re-deriving it. Weaker models stay at the notebook of failures with open guesses. Stronger ones complete the ladder —in the best runs, verifying nearly three out of four learnings—.
The lesson exceeds the models: a loop without distilled memory repeats its mistakes with more energy than anyone. We apply the homemade version here: every failure of this newsroom —an illegible cover, a photo-less publication nobody caught for days— ends up converted into a written rule of the system, not an apology. The recipe is not to avoid failing: it is that no failure repeats.
Where it works, and where it is recklessness
With all of the above, the map draws itself. The loop shines where the result can be measured cheaply, quickly and unambiguously: tests that pass or don't, public rankings, accounting profitability, eval scores. It half-works where measurement is expensive or slow —every iteration of an ads loop costs real money; SEO takes weeks to return signal—. And it is recklessness where the criterion is subjective or the error irreversible: editorial quality, legal decisions, anything touching other people's money or health. There, the correct stop condition has a first and last name: a human. In this house, publishing to the portal is decided by a gate; publishing to X is, additionally, signed by a person. That is not a limitation of the system: it is its design.
What we do not know
Let us declare the uncertainty, as always. There is not yet solid public data on loops running for a year without drift —the claim circulates; the evidence does not—. Nobody knows how much maintenance verifiers demand when the world shifts underneath the metric. And the real cost per outcome —tokens included— varies by orders of magnitude across tasks, with recent studies reminding us that the feeling of speed and measured speed do not always coincide.
The craft being born
What remains is the important part. If generating work becomes too cheap to hold market value, scarcity shifts —again— one layer up: toward whoever defines what «done» means and how it is checked. We wrote recently that the last programming language will be the contract; the loop is its execution engine. Deming dreamed of continuous improvement; tokens have made it executable by the hour and without a union. The future of this craft does not belong to whoever imagines agents working while they sleep —anyone can sell that—: it belongs to whoever can write the stop condition. Because a loop without a serious verifier is not an agency that never sleeps. It is a mistake with an unlimited budget.
