TL;DR
- “Agentic loop” is the phrase of the year, and it is being used for two opposite things.
- In an implicit loop, the AI drives. It decides every step. Wonderful in a chat window, where you are there to catch it. Unpredictable anywhere else.
- In an explicit loop, you drive and the AI rides along, working only where reasoning actually earns its seat.
- Implicit loops make the better demo. Almost every AI product that survives real users is quietly an explicit loop underneath. That gap is where a lot of money is currently being lost.
- Three tools we shipped, three answers: ContentForge (explicit, with a human as the final gate), LeadForge (explicit, model called at exactly two steps), Akigo! Pulse (explicit shell, one bounded reasoning step that admits when it is unsure).
- The only question worth asking anyone selling you an agent: who is driving?
There is a phrase you cannot escape right now if you spend any time near AI: the agentic loop. It is in every pitch deck, every launch post, every conversation about what these tools will do next. And here is the thing I have noticed after building a fair number of them: most of the people saying it could not tell you which kind of loop they mean. They are using one word for two completely different things, and the difference is not a footnote. It is the entire game.
So let me do the unglamorous thing and draw the line clearly, because once you see it you cannot unsee it, and it will change how you judge every AI product someone tries to sell you.
What a loop actually is
An agentic loop is simple to describe. An AI looks at a situation, decides what to do, does it, looks at the result, and repeats until the job is finished or it needs you. That is it. Every AI assistant you have used runs one. So does every tool that claims to be an agent.
The confusion is not about what a loop is. It is about who is driving it. And there are exactly two answers.
In the first kind, the AI drives everything. You hand it a goal and a set of tools and it decides the whole path: what to do first, whether to check its work, when to try again, when to stop. Nothing is hardcoded. You point it at a problem and let it reason its way through. Call this the implicit loop, because the steps live inside the model’s head, not in any code you wrote.
In the second kind, you drive, and the AI rides along. The steps are fixed in advance, written in plain code. The AI still does real work, but only at the specific moments where reasoning is actually the job. The flow from one step to the next is not a decision the model makes. It is a decision you already made. Call this the explicit loop.
Same word. Opposite architectures. One is a brilliant improviser you have to supervise. The other is a machine that does the same reliable thing every time.
Why the difference is the whole story
The implicit loop is what people picture when they hear “agent.” It is flexible, it handles surprises, it recovers from mistakes in creative ways. It is genuinely wonderful in a chat window, where you are sitting there to steer and correct. That supervision is not optional though. It is the price of admission. The same request can take a different path every time you run it. Sometimes it verifies its work. Sometimes it decides to skip that. In a conversation, fine. In a product that a stranger clicks a button on, that unpredictability is a defect, not a feature.
The explicit loop is the opposite trade. It gives up the improvisation and gets back everything a real product needs: the same output from the same input, steps you can test one at a time, a progress bar that actually means something because you know which step it is on, and a cost you can predict before you press go. It is less impressive in a demo and far more useful in someone’s actual week.
Most of the noise about agents is really people showing off implicit loops, because implicit loops make the better demo. Almost every AI product that survives contact with real users is quietly an explicit loop underneath. The gap between those two facts is where a lot of money is currently being lost.
Three things we shipped, three answers to the same question
I would rather show this than argue it, so here are three tools from the Mocoda workshop, each of which had to answer the loop question, and each of which answered it a little differently.
– ContentForge takes one idea and turns it into a month of content. It could have been an implicit loop that reasons about every decision. Instead the workflow is fixed: draft the post, write the version each platform wants, size the images, lay it all out in a folder ready to schedule. The AI is used only where language and judgment live. Everything else is plain code, because resizing an image does not require a mind. And the last step, the one a free-wheeling agent would happily automate, we left to a person on purpose. It never hits publish. The loop is a script, and the human is where it stops.
– LeadForge finds local businesses with weak websites and builds them a finished mockup. Its loop is a pipeline with a set shape: find the prospect, check their site, pull their brand, read their reviews, write a brief, build the page, put it online, track the clicks. Most of those steps never touch a language model at all. Finding businesses is an API call. The color palette is pulled straight out of the pixels of their logo. The screenshot and the walkthrough video are specialized tools that do one job well. The model is called at the two steps that genuinely need it and nowhere else. That is the explicit-loop discipline in one app: spend reasoning only where reasoning is the point.
– Akigo! Pulse is the most interesting of the three, because it is honest about needing both. It is a hub for project and improvement work with more than a hundred defined activities, things like status reports and risk registers. Each activity is an explicit loop: a fixed form with a known output. But inside that form, the AI pre-fills the fields it can infer, which is a small bounded piece of improvisation. The clever part is what happens when it is not sure. A low-confidence guess does not get quietly written in as fact. It surfaces as a question for you to answer. The shell is explicit, the reasoning is contained to one step, and the tool tells you honestly where its confidence runs out. That is the shape almost every good AI product converges on once it has to be trusted.
Three tools. One is explicit end to end with a human as the final gate. One is explicit with the model called at two precise points. One is explicit on the outside with a bounded, honest bit of reasoning on the inside. Notice that not one of them is the pure implicit agent from the demos. That is not an accident. That is what shipping teaches you.
Why I bother making the distinction
None of this is new, which is the part I find reassuring. After a long career running projects, first inside a large financial institution and now in the studio, I have watched a lot of “this changes everything” waves arrive. The ones that actually change things are never the ones with the best demo. They are the ones somebody figured out how to make reliable, boring, and repeatable. The demo shows you what is possible. The explicit loop is what makes it a thing you can hand to another human and walk away from.
So the next time someone tells you they built an agentic loop, ask them the only question that matters: who is driving. If the model decides everything, you have a clever assistant that needs a babysitter. If you decided the steps and the model rides along where it earns its seat, you have something closer to a product. Both are useful. They are just not the same thing, and pretending they are is how people end up trusting a demo with a job that needed a machine.
That distinction is most of what I do now. Not chasing the smartest possible agent, but figuring out exactly where a little intelligence belongs and building plain, honest code around it for everything else. It is less exciting than the pitch deck. It is also the part that actually works.