What It Is

Terminology · LLM Engineering · AI Agents · Benchmarking · AI · ExplainerThe term and its practice, October 2022 to August 2026

• 53days between Cherny's remarks on stage, 15 Apr 2026, and the post that made the idea travel, 7 Jun
• 6accounts of what a loop is, from Cherny's in April to Ng's and Anthropic's on 30 June — at sizes from one prompt to weeks
• 24.69%of full coding tasks finished by the best loop controller in the first benchmark to hold the agent fixed, against 18.52% with none
• 9.4×that controller's estimated cost per run, against the agent left to work alone

Two sayings, one concession, one distinction None of the four is a definition, and that is the point.

• Boris Cherny · 15 Apr 2026

  “My job is to write loops”

  A report of his own practice, not a definition.

  • Said on stage at an evening event run by WorkOS in San Francisco, and published as a video on 2 June. The fullest quotation: “I don't prompt Claude anymore. I have loops that are running. They're the ones that are prompting Claude and figuring out what to do.”
  • His words reach print in two slightly different wordings, and Business Insider reports a third line, given to CNBC: “Claude writes the prompt, and now I'm talking to that new Claude that is kind of coordinating.”
  • The event's host summarises his setup as hundreds of Claude instances reading Twitter, GitHub issues and Slack for product ideas — and adds that most of those ideas are bad today.

• Peter Steinberger · 7 Jun 2026

  “You should be designing loops that prompt your agents”

  The post that made the phrase travel: more than 8.5 million views by September.

  • It opens “here's your monthly reminder” — so by its own account it was not the first time he said it, only the time it caught.
  • Four days later he posted his own loop: tell Codex to maintain your repositories, “wake up every 5 minutes and direct work to threads”.
  • Business Insider describes him as an OpenAI engineer. The post is an instruction, not a definition: it says what to stop doing, and never what a loop is.

• The Claude Code team · 30 Jun 2026

  “You'll come across multiple different answers”

  The vendor concedes the problem, then gives its own answer.

  • Delba de Oliveira and Michael Segner, writing for Anthropic: “If you spend some time on X trying to pin down what a loop actually is, you'll come across multiple different answers.”
  • Their definition: agents repeating cycles of work until a stop condition is met, in four kinds — turn-based, goal-based, time-based and proactive — sorted by what starts a loop and what stops it.
  • And advice against their own enthusiasm: “Not all tasks require complex loops; start with the simplest solution.”

• Armin Ronacher · 23 Jun 2026

  There are two loops

  The distinction that dissolves most of the confusion.

  • “There is already an agent loop inside every coding agent.” The other is “the harness level loop: the loop outside the agent loop” — and “that loop is also not new”.
  • What the outer loop does: it catches the agent as it stops and decides whether that was really the end. The task stays alive past the point where the model by itself would have said “I am done”.
  • His own verdict is uneasy: he has “not had much success with this way of working for code I deeply care about”.

How firm is each of these? Graded first, then told — the standing is in the first column.

Timeline

An old mechanism, a new name Every dated item, oldest first. Only the last is a measurement.

1. 6 Oct 2022The inner loop, as a paper. ReAct has a language model interleave reasoning with actions, so that solving a task becomes a run of thought-action-observation steps. Every coding agent since runs a version of that cycle.
2. 14 Jul 2025An outer loop in one line of shell. Geoffrey Huntley publishes “Ralph”: while :; do cat PROMPT.md | claude-code ; done — the same prompt fed to an agent over and over, one task per pass.
3. late 2025“Closing the loop” catches on, in Andrew Ng's later dating: give an agent a specification and tests and let it iterate. He calls it “a game changer in enabling coding agents to work longer productively without human intervention”.
4. 23 Jan 2026The inner loop, from inside a product. OpenAI's Michael Bolin describes the Codex agent loop, “the core logic… responsible for orchestrating the interaction between the user, the model, and the tools”. A turn ends when the model stops calling tools and answers.
5. 14 Apr 2026Most of the code is around the loop. A study of Claude Code's public source finds its core is “a simple while-loop that calls the model, runs tools, and repeats”, and that most of the code lives in the systems around it: permissions, context management, extensions, sub-agents.
6. 15 Apr 2026The first utterance. At an evening event run by WorkOS in San Francisco, Cherny tells the hosts of the Acquired podcast that he no longer prompts Claude: he writes loops that do. He posts his thanks to the hosts the same night.
7. 2 Jun 2026The recording goes public, with a recap from WorkOS. The same day Steinberger gives a Microsoft Build session, “Build the thing that builds the thing”, on the tools he built to build OpenClaw faster.
8. 7 Jun 2026The post that travels, and the name. Steinberger: “Here's your monthly reminder that you shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents.” The same day Osmani titles a post “Loop Engineering” — and writes, in its first paragraph, “I'm skeptical”.
9. 8 Jun 2026A plain-language account. A Medium write-up of Cherny's remarks gives the most concrete definition collected: a program that prompts the agent, reads what it produced, decides whether the task is done, and prompts again. It dates the remarks to 2 June, the day the video was published.
10. 11 Jun 2026A loop, as its author uses it. Steinberger: tell Codex to maintain your repositories, “wake up every 5 minutes and direct work to threads”, with his own triage and review skills so that “some work can land autonomously”.
11. 20 Jun 2026The mainstream picks it up. Business Insider: “‘Loop engineering’ is all the rage now” — and “the biggest concern cited about loops is, by far, cost.”
12. 23 Jun 2026The two-loop distinction. Armin Ronacher separates the loop inside every agent from the harness loop outside it, and says where he has seen the outer one work: porting code, performance experiments, security scanning, research.
13. 30 Jun 2026Two definitions in one day. The Claude Code team's four kinds of loop; and Andrew Ng's three loops for building products — an agentic coding loop running every few minutes, a developer feedback loop over tens of minutes to hours, and an external feedback loop of users and tests over hours to weeks.
14. 1 Jul 2026The trade press hedges. ADTmag: the term “may prove to be more of a developer shorthand than a formal discipline”.
15. 28 Aug 2026The first measurement. LoopArena, a preprint, holds one coding agent fixed and compares models steering it. On full tasks the best finished 24.69%; the agent left alone, 18.52%; simply restating the goal every round, also 18.52%.

The Argument

Six answers, six sizes They are not competing descriptions of one thing.

What Others Add

What a loop actually does On the mechanism, unlike the definition, the sources agree.

1. 1Put the work where something else can pick it upNot a prompt you type, but a queue, a schedule or an event. Steinberger's own: an agent that wakes every five minutes and directs work to threads. Anthropic's proactive loop runs from an event or a schedule, with no person watching in real time.
2. 2Let the agent attempt it, and stopThis is the inner loop, and it needs no help: the model calls a tool, reads the result, edits a file, runs the tests, and eventually says it is done. Everything that makes it a loop happens after that sentence.
3. 3Disagree with “I am done”The harness decides whether that was really the end. Ronacher's list of what it can do next: continue the same session, inject another message, start a fresh session with modified context, or send the task to another machine. Anthropic's /goal does it with a second model that “checks your condition and sends it back to work”.
4. 4Make the check quantitative, and keep the checker separateAnthropic: “the more quantitative the checks are, the easier it is for Claude to self-verify.” Osmani: split the agent that writes from the one that checks, because “the model that wrote the code is way too nice grading its own homework.”
5. 5Then use the smallest loop that worksFrom the vendor with most to gain from the opposite advice: “not all tasks require complex loops.” And the cost is real: in Business Insider's survey of the discussion, “the biggest concern cited about loops is, by far, cost.”

The first measurement One preprint, one fixed agent, 27 tasks run three times each.

Conclusion

What to take from this Four things, in the order they matter.