Two Ways to Loop an AI Coding Agent, and Why Anthropic Built Both

An engineer finished a $50,000 client contract for $297 in API costs by writing one line of bash: a while true loop that fed a coding agent the same prompt file, over and over, letting it read its own previous work from the filesystem instead of a chat window. Geoffrey Huntley named the technique after Ralph Wiggum, the perpetually confident Simpsons kid, because it works despite being — in Huntley's words — "deterministically bad in an undeterministic world": dumb repetition compensating for whatever the model got wrong last time. In August 2025, a team at a Y Combinator hackathon (a startup accelerator's project sprint) ran the same trick against Claude Code headlessly — no terminal window, no one watching — and woke up to more than 1,000 commits and six fully ported repositories.

By December 2025, Anthropic had turned Huntley's hack into an official plugin. Claude Code also ships a completely different, unrelated command that happens to share half the name: /loop. The two get conflated constantly in writeups and forum threads, and the confusion is understandable — both promise to keep an agent working without a human typing the next prompt. What each one actually does, and when the other is the wrong tool, comes down to a single question: is the goal to keep polling something, or to keep grinding until something becomes true?

Ralph: the loop that lives outside the conversation

The original Ralph technique is almost insultingly simple: while :; do cat PROMPT.md | claude ; done. Every iteration starts a brand-new process with zero conversation memory — the only thing that persists between loops is whatever got written to disk and committed to git. Anthropic's official plugin keeps that spirit but moves the mechanism inside a single session: it installs a Stop hook — a script Claude Code runs automatically whenever a turn tries to end — that intercepts the exit, silently re-feeds the identical prompt, and only lets the session actually stop once Claude outputs an exact, pre-agreed "completion promise" string.

That design has two sharp edges. Running it unattended generally requires the --dangerously-skip-permissions flag, since nobody is around to approve tool calls between iterations. And as the AI infrastructure company HumanLayer points out, Anthropic's plugin captures the surface mechanics of Ralph but not Huntley's actual discipline: carving work into small, independently verifiable chunks with hard stop conditions. Feed it a vague prompt and it will loop just as confidently toward nothing.

/loop: built for checking back, not grinding forward

Claude Code's /loop is a bundled skill — the same first-party category as /code-review and /verify — not a hard-coded built-in, and not a Stop hook at all. It's a scheduler. Give it an interval and a prompt (/loop 5m check if the deployment finished) and Claude Code converts that into a cron job (the standard format for "run this on a repeating calendar interval") and fires it on a fixed cadence. Give it only a prompt (/loop check whether CI passed and address any review comments) and Claude picks its own delay each round instead — anywhere from one minute to one hour — based on what it just observed: short waits while a build is running or a PR is active, long waits once things go quiet. Anthropic's documentation says Claude prints the chosen delay and the reason for it at the end of every iteration.

Under the hood, /loop behaves like a proper scheduler rather than a bash trick: fire times get a randomized offset so every session doesn't hit the API at the same wall-clock second, every scheduled loop hard-expires after seven days regardless of what it's doing, and pressing Esc cancels a pending wakeup instantly. None of that is about reaching a finish line — a fixed-interval /loop will happily poll a dead deployment forever until someone stops it. It's built to check on things, not to complete them.

The command that actually replaced Ralph's job: /goal

Here's the detail that clears up why both exist: Anthropic's own docs put /loop, a Stop hook (Ralph's mechanism), and a third command called /goal side by side in one comparison table, sorted by what starts the next turn and what stops it. /goal is the native version of "keep working until a condition holds." Set one — /goal all tests in test/auth pass and lint is clean — and Claude keeps taking turns until a separate, smaller model (Haiku, by default) reads the transcript after each turn and judges whether the condition is actually met. No exact-string matching, no --dangerously-skip-permissions, no hand-rolled hook script.

That's a strictly more rigorous version of what a completion-promise string is trying to approximate, which is why /goal, not /loop, is the alternative Claude Code's own documentation points to for a Stop hook. /loop was never competing with Ralph in the first place — it occupies the "check back periodically" slot that neither Ralph nor /goal covers.

Which one to reach for

Job	Use
Poll a deploy, babysit a PR, run a periodic maintenance pass	/loop, self-paced
Something on a strict cadence (every 15 minutes, no more, no less)	/loop with a fixed interval
Grind a well-defined task to a verifiable finish line inside one session	/goal
Want the classic overnight brute-force loop, with a hard iteration cap	Anthropic's official Ralph plugin (or the original raw bash version, for CLIs other than Claude Code)
Needs to run with your machine off	Cloud routines (one-hour minimum interval) or a GitHub Actions schedule

The failure mode worth remembering is the same across all five: none of these mechanisms know when a task is a bad fit for looping. /goal's evaluator can only judge what Claude has already surfaced in the conversation, so write conditions its own output can prove — a test-suite exit code, not something taken on faith. And whichever one runs unattended overnight, set an explicit ceiling before walking away: a turn limit on /goal, --max-iterations on a Ralph loop, the seven-day expiry /loop already enforces without being asked. The $297 outcome only makes a good story because it knew when to stop.