Automation
The Loop Is the Easy Part
By John J. BakerJuly 22, 20265 min read
If you have heard someone say an AI agent is "working on it," what they are describing is a loop. The machine tries something, looks at what it produced, decides whether that was any good, changes something, and goes again. It keeps doing that until the work is finished or it runs out of room.
That is the whole idea. It is not complicated, and it is not new. It is also the single biggest reason AI feels different this year than it did two years ago, because a system that gets five attempts and a check beats a system that gets one attempt and a hope.
The trouble is what happens next. Loops are exciting, and the excitement makes people stop early. A loop will carry you a long way, and then it will stop dead at exactly the point where the work starts to matter.
What a loop actually is
Before anything else, here is the shape of it, in four beats.
Act
It does the thing.
Drafts the email, prices the line item, pulls the numbers together. One attempt, start to finish.
Observe
It looks at what came out.
Reads its own work back, or checks the result against the record. This is the step that separates a loop from a one-shot answer.
Judge
It decides whether that was any good.
Better or worse than the last attempt, and against what standard. This is where nearly every loop is thin, because the standard was never written down.
The weak beatAdjust
It changes something and goes again.
A different angle, a missing input, a narrower scope. Then back to the top, until it is good enough or it runs out of room.
Beat four hands back to beat one, and the whole thing runs again. That repetition is what people mean when they say an agent is working on something.
The tool gives you beats one, two, and four for free. Beat three is the one you have to supply, and it sets the ceiling on everything the loop can reach.
Notice which of those four you get for free. The tool will act, it will look at what it produced, and it will change something and try again. Those are mechanical. Every agent product on the market does them.
The third beat is the one nobody hands you. Judging whether an attempt was good requires knowing what good looks like here, in this business, for this customer, on this job. That standard exists. It lives in your people's heads and it has almost certainly never been written down.
Why loops get you most of the way there
I want to be fair to the loop, because the case against stopping early is not a case against starting.
A loop is genuinely the right foundation. It turns a system that produces one confident answer into a system that produces a draft, notices what is missing, goes and gets it, and produces a better draft. It handles the drafting, the gathering, the arithmetic, the reformatting, the first pass at almost anything repeatable. If you are not running loops yet, that is where to start, and the gain is real on day one.
Where the reasoning goes wrong is in the extrapolation. The loop covers a large share of the work quickly, so it looks like a few more turns of the same crank will cover the rest. It will not, because what is left is a different kind of problem.
What the loop cannot reach
What the loop reaches, and what it does not
- Most of it. The loop handles this on its own.
- Drafting, gathering, checking its own arithmetic, a competent first pass at almost anything repeatable. Genuinely useful, and genuinely most of the work.
- The last slice. It decides whether any of it is usable.
- Not harder work. Unwritten work. The judgment your people apply without noticing they are applying it.
This one gets a phone call, not an email
Because of how the last job ended, which is in nobody's notes and everybody's memory.
That number is low and the photos say why
Access is bad on that site. The takeoff was right and the estimate is still wrong.
We do not quote that client the list price
There is a history, a reason, and a person who would be insulted by the standard letter.
Stop, this one needs the owner
The signal that a job has turned into a relationship problem instead of a scheduling one.
Good enough to send, today
The standard moves with the client, the season, and how the week has gone. It is real, and it is not written anywhere.
Look at what is actually in that last slice. None of it is hard in the sense of being complicated. It is hard in the sense of being unwritten.
Your estimator does not consult a rule when he looks at a takeoff and says the number is light. He looks at the site photos, notices the access is bad, and knows. Your best rep does not run a decision tree before deciding this one gets a phone call instead of an email. He remembers how the last job ended.
We call that intuition, and the word does us a disservice, because it makes the knowledge sound mystical. It is not. It is compressed experience. It is a thousand small corrections a person has absorbed over fifteen years, filed somewhere they cannot easily reach and cannot easily explain. Ask them to write down how they price a difficult client and you will get half a page that misses the thing they actually do.
The machine has none of that. It has read a great deal about construction and nothing at all about your company. It will produce work that is correct in general and wrong here, and the gap between those two is the entire distance between a demo and a system your team will actually use.
A loop that cannot tell better from worse is not iterating
This is where I would push back on the way loops usually get talked about, including by people selling them.
A loop with no standard is not iterating. It is repeating. It will run its ten turns, produce ten variations, pick one, and hand it to you with complete confidence, and none of those turns made it better because nothing in the system could tell better from worse. Motion is not progress. More turns of a loop with no judgment in it just gets you to the wrong answer faster and with more supporting detail.
So when people say the winners will be whoever iterates their loops best, I would sharpen it. The loop is not the advantage. Loops are commodity, they ship in the box, your competitor has the same ones. The advantage is the standard you feed into that third beat, because that standard is built out of your specific history and nobody else can copy it.
Which is a more encouraging conclusion than it sounds. It means the durable edge here is not technical skill. It is knowing your own process well enough to describe it.
How the last slice actually closes
The good news is that this knowledge is recoverable, and the recovery happens in a very ordinary place: the moment a person overrides the machine.
The override
A person changed it before it went out.
The rep rewrote the email. The estimator moved the number. The work got done and the machine learned nothing.
The reason
Why did they change it?
Not what they changed, why. One sentence, captured at the moment of the edit, while the reason is still obvious to the person making it.
The rule
Is this a one-off or a pattern?
Five reasons that rhyme are a rule. Naming it turns a single correction into something that applies to every case like it.
The system
Where does the rule now live?
In the instructions, in the examples the model reads, or in code if it has a right answer. It stops being tribal knowledge and starts being infrastructure.
A loop only gets better if something outside the loop is keeping score.
Left alone, a loop repeats. It becomes iterative the moment human corrections stop disappearing and start being written down as reasons.
Every override is a data point about the standard. The rep rewrote the email, so the draft was wrong in a way he could feel. The estimator moved the number, so the model was missing something he could see. That correction contains exactly the judgment we said was unwritten, and it is available for about thirty seconds before the person moves on and forgets they made it.
Most teams capture the fix and lose the reason. The email goes out, corrected, and the same correction gets made again on Thursday. The person doing it stops noticing they are doing it, which is how a team ends up quietly working around a tool that everyone still describes as helpful.
The discipline is to capture the reason instead. One sentence, at the moment of the edit. Do that for a month and the reasons start to rhyme. Five reasons that rhyme are a rule, and once you can name the rule you can put it somewhere it will run every time: into the instructions the model works from, into the examples it reads before it starts, or into plain code when the question turns out to have a right answer. That last case is more common than people expect, and it is the difference between a system that surprises your team and one that can explain itself.
That is what people are gesturing at when they talk about improving how an AI thinks over time. It is not a clever prompt. It is a slow accumulation of your judgment, one override at a time, into a place where the loop can reach it.
Some of that last slice should stay human, permanently
Now the honest caveat, because the arithmetic of "get to one hundred percent" is the wrong goal.
Some of what is in that final slice should never be automated, no matter how well you encode it. The call where you tell a client the schedule slipped. The decision to walk away from a job. The judgment that a relationship has gone sideways and needs the owner rather than another follow-up. Those are not gaps in the system. They are the things worth keeping in human hands even on the day the machine could technically handle them.
So the target is not full autonomy. The target is a system that handles the bulk of the work well, knows where its own edge is, and stops there instead of guessing. A loop that says "this one needs you" is worth more than a loop that produces something plausible and lets you find out later. That, in practice, is what a mature loop looks like: not one that never stops, but one that stops in the right places.
And the second honest caveat: the loop is not what makes this work. An agent with no clearly named job will loop enthusiastically toward nothing in particular. Before any of this, somebody has to be able to say what the machine is for and how you would know on Friday whether it helped.
Where ClearOak comes in
Nearly everything we do sits in that last slice.
Anyone can stand up a loop now. That part has genuinely gotten easy, and if that is all you need, you do not need us. What is still hard, and what we spend most of our time on, is sitting with the estimator until we understand why he knew the number was light. Then working out which part of that is a rule, which part is a judgment call the model can be taught, and which part should stay with him permanently.
That is unglamorous work. It looks like watching people do their jobs and asking why more times than is comfortable. It is also the only part that compounds, because every piece of it you capture is a correction nobody has to make again.
If you have an AI tool that produces work your team keeps quietly fixing, the fixing is the signal. Those corrections are your standard, and right now they are evaporating.
Schedule a call and tell me what your people keep correcting. That conversation usually finds the last slice within twenty minutes.
Prefer email? Reach me directly at john@clearoakconsulting.com and describe one output your team always edits before it goes out. That edit is the whole thing.