The Ralph Wiggum loop, and what it actually costs when a client is paying

TL;DR. In late 2025 the Ralph Wiggum loop went viral: a wonderfully simple way to run a coding agent unattended, in a loop, until it finishes. The internet's reaction was mostly "cute trick" and a wave of "I left it running overnight and woke up to a thousand commits" screenshots. We did something less glamorous. We ran the idea against real client work, watched where it cost money and where it broke, and kept the good part. This is what the Ralph loop actually is, what it costs when a euro is attached, and what it became in our hands.

The hype, and the good idea underneath it

Geoffrey Huntley published the Ralph Wiggum technique in 2025, and by the final weeks of the year it was everywhere. Credit where it is due: it is his idea, and it is a good one. The name is self-deprecating, after the Simpsons character who is not the sharpest, because the technique is almost embarrassingly simple. In its purest form it is a loop that runs a coding agent again and again, each pass starting with a fresh context, feeding the previous result and its errors back in, until the thing works.

The clever part is the fresh context. Long agent sessions rot: the more an agent piles into its working memory, the more it loses the thread, contradicts itself, and drifts. Starting each pass clean sidesteps that. Instead of one long confused conversation, you get many short sharp ones, each building on the last one's committed output. That is a genuine insight, and it is why the technique stuck around after the screenshots stopped.

So the loop is good. The problem was never the loop. The problem is the word next to it in every viral post: "overnight." Unattended.

What "unattended" costs when a client is paying

The overnight-coding genre works as a demo because the stakes are zero. It is a personal project, a weekend, nobody is charged if it produces garbage. Attach a client and a production system, and "unattended" changes from a fun adjective to a financial and reputational risk. Here is what we mean, from our own runs, not a survey.

We have watched an agent complete a task, pass 239 tests, and deliver nothing that worked, because a real-world detail broke silently in a way no test caught. Green the whole way down, and useless. On a personal project you shrug. On a client's clock, you paid real money for an agent to confidently produce zero, and without a human looking, you would not have known until it mattered.

We have also watched a spend cap fail. We set a limit of one euro a day on an autonomous process, sensible and cautious. It sailed past to nearly double that, because the check looked at money already spent rather than the worst case still committed. A small number, deliberately, because we were testing. Scale that logic to a real budget running unattended overnight and the small number is not small anymore.

Neither of these is a reason to fear the technology. They are the reason the "unattended" part is a marketing position, not an engineering one, the moment real money is involved. The technology is cheap. An unwatched mistake is not.

Ralph on the Oversight Ladder

The cleanest way to place the Ralph loop is on the Oversight Ladder, our five levels of AI coding autonomy. Ralph, run the viral way, is Level 5: fully autonomous, unattended, a person reviewing the batch afterward if at all. Level 5 is the right setting for exactly one kind of work, throwaway exploration where a silent mistake costs nothing. It is the wrong setting for anything a business runs on.

The loop is identical. The oversight is not. That single column is the whole difference between a viral screenshot and something we will run on a client's business.

What it became for us

We kept the fresh-context loop and threw away the "overnight." In practice that means the loop runs, but inside guardrails: automated tests and checks gate every pass, and a senior person can see and stop the run. Agents cannot merge their own work. The loop still gives us its real gift, short sharp iterations that do not rot, but a mistake meets a wall before it meets a customer.

That is the honest arc of most good techniques. They arrive as a flashy demo, the demo is not the useful part, and what survives contact with real work is the quiet mechanism underneath. Ralph gave the industry a genuinely better loop. What it did not give anyone, and what nobody's overnight screenshot shows, is the oversight that makes it safe to point at something that matters. That part you have to build, and it is most of the job.

The question to ask about any "autonomous" pitch

If a vendor sells you "autonomous" AI development, the Ralph hype taught the useful question: autonomous, and then who looks? If the answer is that the AI is so good it rarely errs, you are being sold the overnight screenshot. If the answer is that the loop runs inside gates a person controls, and a failing check stops it before anything ships, you are being sold an engineering practice. The first is a hope. The second is what you actually want to pay for. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, largely for cost, unclear value, and weak risk controls. Most of that is the overnight screenshot meeting a real budget.

We build production software with agent loops every day, with the oversight this article describes. If you want to know which of your work could safely take more autonomy and which could not, that is what a first conversation is for.