Most marketing teams still use AI one turn at a time. Write a prompt, read the answer, fix the brief, ask again. The tool is fast. The person is still the scheduler, the reviewer and the memory for every step.
Loop engineering moves that work into a system that repeats. The system picks up a task, gives the agent the right material, checks the result and decides whether to go round again. The human sets the boundaries and takes the calls that need judgement.
The phrase came out of software teams, and the working idea travels. A weekly content programme already repeats. What changes is writing down how it repeats, including what counts as good enough and when the thing has to stop.
Where the phrase came from
In a 7 June 2026 account, Addy Osmani quotes Claude Code creator Boris Cherny saying he had loops prompting Claude and deciding what to do. No searchable first-party transcript turned up, so read that as Osmani's dated transcription of the remark.
Peter Steinberger put it more bluntly in a 7 June post, telling developers to design loops that prompt their agents. Osmani published "Loop Engineering" the same day, gave the practice a name and described the parts around it.
The term is still unsettled. Osmani helped popularise it, but there is no good evidence he used it first. People also mean different things by "loop": a scheduled prompt, an agent checking its own work, a bigger system running several agents. Treat the label as useful shorthand, not a standard.
The progression is cumulative
The progression is prompt > context > agent operating setup > loop. A prompt says what to do in one turn. Context supplies the facts, examples, brand rules and prior work. The operating setup grants approved tools, file access, permissions and limits. The loop adds recurrence: it decides when work starts, passes in the current context, asks for an action, checks the output, records what happened, then repeats or stops.
Every layer sits inside the next one. A loop still needs a precise prompt. Thin context still produces thin work. Sloppy permissions can expose the wrong files, or send a draft further than you meant. What a loop changes is where your attention goes, away from typing each instruction and towards the operating rules.
A June 2026 position paper draws the same cumulative distinction. It defines a loop as a bounded specification with a trigger, a goal, a verification step, a stopping rule and stored state. That definition helps marketers, because it separates a loop from automation that fires again regardless of how the last run went.
Build one weekly content loop
Pick one channel and one recurring job. A weekly LinkedIn content loop is small enough to inspect and frequent enough to justify the setup. Write its five parts before you connect a single tool.
- Goal. One publishable post for an agreed audience and topic, ready by Thursday afternoon.
- Context. Approved content themes, the voice guide, recent posts, the source policy, current campaign notes.
- Small action. Select one topic, research it, draft one post. Keep publishing outside the loop at first.
- Check. Test source dates, unsupported claims, repeated angles, voice rules and the channel's length limit. Failed checks go back for one revision.
- Stop rule. Stop after approval, after the revision limit, or when a source or brand decision needs a person.
State can live in a plain content tracker: topic chosen, sources used, check results, revision count, final decision. Next week's run can then see what has already been covered. Without that record, the system repeats an angle or rediscovers the same facts every Monday.
Keep the human steering the work
A loop earns its place only when the check can reject weak work. That is the whole test. A vague instruction to keep improving will run for many turns, spend tokens, and still hand back an ordinary draft. More agents also mean more model calls, more context sent, and more material somebody has to read.
A 26 June practitioner field note puts the cost objection plainly: an unattended loop can burn tokens without producing better work. Topic choice, risky claims, brand judgement and publication stay with a person. Set a run limit and a spending limit. Keep the first loop in draft mode and review its failures for a month before it is allowed to touch anything external.
Pick one weekly task and write its goal, context, action, check and stop rule on a single page. If you cannot state the check plainly, keep using a direct prompt until you can.