Loop Engineering

Teaching an AI to check on its own, no prompt needed

Loop engineering means setting up an AI so it checks on something by itself, again and again, on its own schedule, instead of waiting for a person to ask each time. It builds on top of simpler ideas like writing a clear instruction or giving the AI a plan for a big job. For a freelancer or VA, it matters because it shows where AI tools are heading: less 'please do this now' and more 'keep watching this for me.'

What loop engineering means

Loop engineering means setting up an AI so it checks on something by itself, over and over, on its own schedule. Instead of a person asking the AI to do a task each time, the AI starts the task on its own. The AI looks at a situation, decides what to do, does it, checks the result against a rule, and repeats. It only stops when that rule says the job is finished. This repeating pattern is why it is called a loop.

The problem it solves

To see why loop engineering came up, it helps to look at the ideas that came before it. Each one was invented because the one before it hit a limit.

Prompt engineering is writing a clear instruction. It works well for questions the AI can already answer using what it knows, like a simple math question.
Context engineering lets the AI go get extra facts on its own, such as searching the internet, before it answers a harder question. This works well, but it struggles once a job takes longer than a few minutes.
Harness engineering was built for long jobs. A long job needs more information than the AI can hold at once. If the AI keeps shortening its own notes to make room, important details start to disappear. Harness engineering is an outside system that tracks the whole job and breaks it into smaller steps, so the AI can work through a long list of tasks without losing the plot.

Loop engineering stacks one more layer on top of harness engineering. It does not remove the need for a clear instruction, useful facts, or a solid plan. What it changes is who starts the work. In every stage before this one, a person had to ask the AI to begin. In loop engineering, the AI begins on its own, based on a schedule or a rule that was set up in advance.

A circular diagram showing an AI reasoning about a task, acting on it, checking the result against a rule, and repeating until the rule says the job is done — The same four steps repeat: reason, act, check, and go again, until the rule says stop.Open full size

How the loop actually runs

Reason. The AI looks at the situation and decides what needs to happen next.
Act. The AI does one piece of work, like updating a number or trying a fix.
Check. The AI compares the result against a clear, checkable rule, not a feeling.
Repeat or stop. If the rule is not met yet, the AI goes back to reasoning. If the rule is met, the loop stops.

What a loop needs to actually work

A schedule or trigger. Something that starts the check on its own, like 'every hour', with no person clicking a button.
A safe, separate work area. So that two checks or fixes running at the same time do not bump into each other or overwrite each other's work.
Saved knowledge and tools. Things the AI already knows how to use, so it does not have to start from zero every single time it runs.
A checker. A second AI, or a test, that reviews the first AI's result before it counts as truly done.
A memory of where things stand. A record of what has already been checked, fixed, or is still waiting, so the loop does not repeat work it already did.

Who checks the work

The Check step is where most loops quietly fail, because the thing checking is usually the thing that did the work. An AI asked whether it did well says yes. It is not lying. It cannot see the mistake it just made, for the same reason it made it.

On the left one worker doing the job and then grading its own work, marked as marking its own homework; on the right a separate rule that passes or sends it back — The checker can be dull. What it cannot be is the same worker in the same run.Open full size

A real checker is separate, and it can be very boring: a test that runs, a number compared to a target, a second agent applying one written rule. What matters is that the rule existed before the work did, so it cannot be quietly reinterpreted to fit whatever came out.

Four brakes every loop needs

Four panels: a try limit, a spending cap, an approval gate for anything irreversible, and a log you can read afterwards — Four brakes. Decide all four before the loop runs once.Open full size

A try limit. After five goes, stop and message a person. Something it cannot solve is something it will retry forever.
A spending cap. A hard limit on the provider key. A loop running overnight can send thousands of messages while you sleep, and each one is billed.
An approval gate. Anything it cannot undo waits for a yes: sending, paying, deleting, publishing.
A log you can read. What it checked, what it changed and when. Without one you cannot tell a loop that is working from a loop that is stuck.

A real example you can picture

One builder set up an AI to run a website that shows live sports scores during a tournament. Scores change many times a day, so updating the site by asking the AI each time would mean typing the same request over and over. Instead, the builder set up a loop: check for new scores every hour, and update the site automatically when something changed. A second loop watched for bug reports from visitors and tried to fix small issues on its own. Neither loop needed a person to press go. Each one followed its own schedule and its own rule for when a fix counted as finished.

When to use a loop, and when not to

Good fit for a loop	Not a good fit for a loop
A repeating check with a clear rule, like 'is this price still correct'	A one-time task you only ever need once, like writing a single email
Small, low-risk fixes that are easy to undo	A decision that needs real human judgment, like which client to prioritize this week
Something you would otherwise have to remember to check yourself	Anything that spends real money or deletes real data without a person confirming first

Mistakes that break a loop

A vague stop rule. 'Keep going until it looks good' is not something a computer can check. 'Stop when the test passes' is.
No limit on repeats. A loop that can run forever can also waste time and money forever. Always set a maximum number of tries before it must stop and ask a person.
No one checking the risky steps. A loop is great at small, repeatable checks. It should still stop and ask a person before doing anything that is hard to undo.
Treating it as already proven. This idea is still new and untested at scale. Some people who build with AI every day call it mostly hype so far. Try it first on a small, safe task, not your most important system.

More words

Prompt Engineering
Writing the instruction so the answer comes back right
Context Engineering
Choosing what the AI gets to see while it answers
Graph Engineering
How many AI agents connect and work as a team.

See all 11 explanations

Reading time: 10 min
Group: How agents are built

Words inside this word

Prompt engineering: Writing a clear instruction that tells an AI what you want it to do.
Context: The information an AI can use right now, like the papers open on your desk at this moment.
Context engineering: Letting an AI go get its own extra facts, like searching the web, before it answers something.
Harness engineering: An outside system that manages a big job for an AI, breaking it into smaller steps so nothing important gets lost.
Loop engineering: A system that lets an AI start its own checks and tasks on a schedule, without a person asking each time.
Automation: Work that repeats by itself, on a timer or a trigger, with no person clicking start each time.
Sub agent: A second AI that checks or tests the first AI's work, like a proofreader checking a writer's draft.
Stop rule: The exact, checkable condition that tells a loop when to stop repeating.
Brake: Any limit that can stop a loop: a cap on tries, a spending cap, or a step that waits for a person.
Approval gate: A point where the loop stops and waits for a human yes before doing something it cannot undo.

Where this came from

Watch the original video