From Prompting to Directing

Almost everyone I talk to uses AI the same way, and for a long time I did too. You open a chat, you ask it something, you read what comes back, you use it or you do not, and you close the tab. It is a good assistant, a fast writer, a search that talks back. That is a real and useful thing. It is also the near edge of what it can do, and it is easy to stay there without knowing there is more further in.

The last piece was about the first step past that. It made the case that a loop, the thing the people building these tools have started organising their work around, is a task set up to run itself. You write it once, and it goes, without you sitting in the middle of every turn. That is already a different relationship with the machine than asking it questions. It is the difference between using a tool and directing one.

This piece is about what happens after that first step. A loop is the first move away from the chat window, and there are two more past it, usually called orchestration and graph. They are not jargon to memorise. Each one is a way of working that is larger than the last, and further from where you started. The names are ordinary enough. What they let you do is not.

Start where the last piece left off. A loop is a task that runs itself, the weekly report or the standing chase, set going without a person kicking off each turn.

That much sounds like automation, which is an old idea, and if that were all it was it would not be worth a second piece. Here is the part that is not old. A task a person does happens once and leaves nothing behind. They do it, they move on, and next week they start again from the same place. A loop does not start from the same place. It keeps what worked, notices where it went wrong, and runs the next time a little better than the last. Over a hundred runs it is not doing the same task a hundred times. It is getting good at the task, on its own, in a way a person repeating a chore never does.

So the shift is easy to state and takes a while to absorb. You are not automating a task. You are setting something running that improves at the thing while you are not watching. The person who wrote it is not doing the work and is not repeating it either. They are further up, deciding what the loop should be aiming at and whether what it produced can be trusted. That is a different altitude to work from, and it takes a little while to get used to standing there.

Once you have more than one loop, something has to decide what runs when, and in what order, and what waits for what. The research has to finish before the summary. The numbers have to be checked before they go in the report. That deciding is what orchestration means.

In a company that job has always existed. It is the person who kept the work moving between other people, who knew that this had to happen before that, who chased the piece that was late. Coordinating. Every organisation runs on it.

The reason this matters is not that a machine can now do that coordinating. It is what the coordinating stops limiting. A person keeping the work moving can hold a handful of threads at once, and only while they are at their desk. That limit has shaped every organisation ever built, mostly without anyone noticing. The reason work comes in departments, in handoffs, in things that take three weeks, is partly that a human can only hold so much in flight at one time. When the coordination is written down and runs on its own, that ceiling lifts. A business can suddenly run far more moving parts at once than it could ever have staffed, which means it can take on work that used to be waved away as too big, too fiddly, or not worth the people it would tie up. The value is not a tidier process. It is that things which were never worth attempting come into reach.

That is the second rung. The first was a task that improves itself. The second is a business that can hold more at once than its people ever could. That is already a long way from asking a chatbot a question.

The third move is the largest, and the one still finding its feet.

Everything in a business is connected to everything else. This customer sits under that contract, which carries this risk, which traces back to a decision someone made two years ago for a reason only they remember. Nobody holds all of those connections. The person in finance holds some, the person in sales holds others, the one who has been there twenty years holds the threads nobody wrote down. The knowledge of how the company actually fits together is real, and it is scattered across a lot of heads, and no single person has ever held the whole of it at once.

A graph, in the sense people are reaching for now, is that web of connections written down in a form a machine can follow. Not a folder of documents to search through, but the relationships themselves, mapped, so the thing can move along them the way the twenty year employee does in their head when you ask them a question in the corridor.

When the connections are written down, you can ask your own business a question that no one person in it could answer. Not because the answer was hidden, but because answering it needed sales and finance and legal and history all held in one mind at the same moment, and no one mind ever held them. Which of our customers look like the ones we lost last year, and why. Where are we carrying a risk that only shows up when you put the contract and the support history and the invoices side by side. These are not searches. They are questions about how the whole business hangs together, and until now the only way to answer them was to get the right six people in a room and hope someone remembered.

That is the height these three moves have been climbing towards. Not a faster assistant. Not even a tireless one. Something you can ask about your own company and get back an answer that draws on more than any individual ever carried. The chat window you started in asked the machine what it knew about the world. This asks it what your business knows about itself.

There is a practical side to all this, beyond the capability, and it is where the last piece and this one meet.

The models are becoming the same for everyone. The one you use and the one your competitor uses are converging, and the gap between them is closing fast. The loops matter, but a competitor can build the same loops from the same models. What they cannot build is your version of that map, the connections particular to your business, your customers, your history, the decisions and the context that exist nowhere else. Point a capable model at the open web and you learn what is happening in the world, the same as everyone else. Point it at your own map as well, and it can tell you what any of it means for you. That second half is the part no one can buy and no vendor can sell.

The clearest public account of a company living this way came recently from Amjad Masad, who runs Replit. He described what has happened inside his own company over six months, and it reads like a different sort of organisation, one where the loops run and the people direct them. I would read it with some caution, because it is his account of his own company and he sells the infrastructure the story is advertising, and the headline number he leads with, lines of code, is close to meaningless as a measure of anything, which he knows better than I do. But look past the figures to the shape of it, and the telling detail is not in any of the charts.

Before any of it worked, his team had to give the system a map of the company’s own information. Which numbers are the real ones, which to trust, how it all joins up. Every company has sat in the meeting where two people have different figures for the same month and no one can say who is right. It rarely gets fixed, because you cannot get a budget for tidying up and it loses to the next launch every time. That unglamorous, endlessly deferred work turns out to be the ground everything else stands on. The clever part was never the hard part. The map was.

So these are worth understanding, but not as vocabulary. Loop, orchestration, graph, whatever they end up being called, are three steps up and away from the chat window, towards a way of working that is still new to almost all of us. The move is not to use a better tool. It is to stop thinking of it as a tool you talk to, and start thinking of it as something you point at your own business and direct. That is not a small change in speed. It is a change in the kind of work that becomes possible, and it is open to anyone ready to see the machine differently.

If this was useful, subscribe. This follows an earlier piece, The Person Was the Machine, on why the person was always the loop and what changes when the loop can run on something else.

Craig Hepburn is an AI strategist and Perplexity Fellow. Twenty years building at the frontier of digital, from Microsoft and Nokia to Art Basel and UEFA. Now building at the frontier of agentic intelligence.