From prompts to loops: the next evolution of working with AI

TL;DR

We started with prompts. Then we needed context, so context engineering arrived. But the piece we kept missing was never the journey. It was the intent. Agents and models got good enough to work out the how on their own. So the new unit of work is the loop: you tell the agent to keep going until it hits the outcome you defined, not until it spits out an answer. The skill was never clever prompting. It is being brutally clear on the destination, connecting the right context, and getting out of the way.

How did we get from prompts to loops?

Three years ago the whole game was the prompt. Type the right words, get the right answer. We collected prompt templates like recipes.

Then the cracks showed. The model didn't know your business. It didn't know your data, your tone, your standards, your last six months of decisions. So we got smarter about feeding it context. We called it context engineering, and it was a real step up. Suddenly the AI could answer as if it actually knew where it worked.

But the work still missed. Not because the model was thick. Because we were vague about what we actually wanted.

Meanwhile, two things were quietly changing. The agents got better at acting, not just answering. And the models got better at reasoning their way through a messy task. Put those together and you get the loop.

That is the arc. Prompts gave us answers. Context gave the answers a home. Loops give us outcomes.

Why did prompts stop being enough?

A prompt is a single instruction with no memory.

Ask it cold and you get generic work, because it is answering for everyone, not for you. Garbage in, generic out.

Context engineering fixed a chunk of that. You feed the model your documents, your data, your examples, the guardrails it has to stay inside. Now it answers from inside your business, not from the open internet. This is where most serious AI work lives today, and it matters. An agent that knows your numbers, your customers and your rules is worth ten that don't.

So we had context. And the output still wasn't right. Why?

If we had context, why did the output still miss?

Because we loaded everything the agent needed to know, and stayed fuzzy on what we actually wanted.

We told it about the business. We never told it, precisely, where we wanted to end up. We hoped it would fill in the gap. And it did, with its best guess, which is almost never your guess.

Let me be clear about what I mean by the gap. I do not mean the journey. I do not mean the steps, the method, the how. I mean the intent. The outcome. The goal. The destination. That is the thing we keep leaving vague, and that is the one thing the agent cannot read your mind about.

The agent can find the road. It cannot guess where you wanted to go.

What is a loop, exactly?

A loop is simple.

You tell the agent to keep working, checking its own output against the goal, and going again, until it actually achieves the outcome. Not until it produces something. Until it produces the thing you asked for.

The old way was one prompt, one output. You read it, you judge it, you accept it or you start again from scratch. All the judgement sat with you.

The loop way moves the judgement into the work. The agent drafts, checks it against your stated outcome, finds the gap, fixes it, checks again. It keeps going until done means done.

This is exactly why agents felt almost good enough for so long. They produced an output when you wanted an outcome. The loop is what closes that last gap. It is the difference between a clever assistant and one that finishes the job.