Agent loops: the end of prompting AI

The person who built one of the most popular AI tools for developers in the world recently said something that reads like clickbait:

"I don't prompt AI anymore."

It isn't clickbait. That's Boris Cherny, head of Claude Code at Anthropic. And the full quote is even more interesting:

"I don't prompt Claude anymore. I have loops running. They prompt Claude and figure out what to do. My job is to write the loops."

A month ago that talk went viral. Good — because it shows something most business owners are still missing. While you're busy learning to "write better prompts for ChatGPT," the best people already stopped playing that game entirely.

The new game is called agent loops. I'll explain it plainly, in human terms, no tech jargon — what it is, why it blew up right now, and what it can actually do for your business.

Three levels of using AI

To see why this is a real shift, picture three levels.

Level 1 — conversation. You open ChatGPT, ask a question, get an answer. This is where 95% of people are. AI is a consultant you ask for advice. The problem: you have to ask every question, check every answer, and stitch it all together by hand. The AI does nothing on its own. It waits for you.

Level 2 — the agent. Instead of just talking, the AI gets tools: it can search the web, open a file, send an email, check a calendar. And it gets a goal. "Find me 20 property developers in the city and put them in a spreadsheet." The agent decides which tools to use, and in what order. That's an employee, not a consultant. But you're still sitting over its shoulder, watching.

Level 3 — the loop. This is where the jump happens. Instead of launching the agent by hand every time, you build a system that launches it for you — over and over, at set times, or until the job is done. You don't prompt. You design the machine that prompts for you.

Cherny put the jump like this: "Going from agents to loops is as big a jump as going from code to agents."

It's the difference between "I have a tool I use" and "I have an employee who gets the work done on their own — including while I sleep." At Level 2 you're the operator. At Level 3 you're the manager. And a manager who has to redo every task by hand isn't really managing anything.

What an agent loop actually is — no tech

Picture an intern you hand a sheet of instructions:

1. Check what needs doing.
2. Do the next step.
3. Check whether it came out right.
4. If not — fix it.
5. Go back to step 1. Repeat until everything's done.

An agent loop is exactly that — except the intern is the AI, and the instruction sheet is code that doesn't get tired and doesn't take breaks. The loop spins on its own: the AI does a step, checks the result, fixes the error, does the next step. Until it hits the condition "done."

And here's the part you need to hear: this is not science fiction, and it's not tech reserved for corporations with an IT department. If you already use modern AI tools, this is probably within reach for you right now. What changes is the way you think — from "what do I ask the AI to do today" to "what system do I build so the AI does it without me."

Why now? Three things came together

Because the idea of "AI in a loop" isn't new. What's new is that it finally works. Three things lined up.

First, the models finally use tools well. A year ago, an AI turned loose in a loop would lose the thread after a few steps. Today it can open a file, run something, see an error, understand it, and dig itself out — without you.

Second, the tools matured. There are now ready-made ways to run these loops — locally on your own computer, or in the cloud, where they keep going even after you close the laptop. You don't have to build them from scratch.

Third, someone showed it live. Cherny described how engineers at Anthropic actually work — and the company openly writes about how it builds such systems. Not "in theory it's possible," but "this is what we do every day." And at a scale that's hard to ignore: dozens, even hundreds of agents at once, some working through the night. One example that made the rounds — a single massive change broken up into hundreds of agents working in parallel across thousands of files, each handling its own slice. Something a human wouldn't finish by hand in any reasonable time.

The key sentence I take from all this, for you: the edge no longer sits in how smart the model is. It sits in how well you design the system around it.

What this actually gives a small business

Let's leave Silicon Valley. What does a one-person agency, a clinic, a law firm, or a shop get out of this?

Think about all the things you do on a cycle that eat your time even though they're repetitive:

• Monitoring — the daily check for whether anything important came up. A new tender. A move by a competitor. A mention of your company. A loop can do this every morning and hand you a finished summary before your coffee's gone.
• Reports — the same roll-ups every week. Sales, traffic, campaign results. Instead of typing them out by hand, you wake up to a finished report in your inbox.
• Research — background on a client before a meeting, market analysis, a scan of what's happening in your industry. A loop gathers, summarizes, and leaves you a note.
• Content — and this is my favorite example, because I'm building exactly this for myself right now (more on that in a second).

Take a concrete one. Say you spend the first half-hour of every Monday pulling last week's numbers into a report your team actually reads. Same sources, same layout, same questions, every single week. That's not a decision — it's a chore. It's a loop waiting to happen. The half-hour you get back isn't the point; the point is you stop being the person the report depends on.

The rule is simple: if you do something regularly, on a repeatable pattern — it's a candidate for a loop. You don't have to sit there for it. You have to design it well, once.

Not sure which of your processes is a fit for a loop? Book a 30-min call — we'll walk through it together.

How I'm building this for myself — the content example

I like specifics, so here's one. I'm building a loop at Prospere right now that's meant to do what currently costs me hours every week: track AI news and viral content, spot the best topics, and write first drafts of posts from them.

The loop's logic goes like this:

1. Every day it searches the sources I already follow — X, industry news, whatever's getting clicks.
2. It judges what's genuinely worth attention and what's just noise.
3. From the best topics, it writes first drafts — for the blog and for LinkedIn.
4. It leaves them for me to review, in one place.

In the morning I no longer start from "what do I write about today?" I start from finished drafts I only polish and publish. The machine does 80% of the work — the digging, the selection, the first draft. I do the 20% that needs a human — judgment, voice, the call on what to ship.

That's exactly what Cherny meant. You stop being the person who sits and prompts. You become the person who designs and watches the system.

Honestly: this is not magic

Now the part you won't hear from the hype sellers on Instagram.

A loop is not a fix-all. Most business automation isn't autonomous loops at all — it's plain, predictable processes: "when X happens, do Y." And that's a good thing, because those processes are cheaper, more stable, and easier to stand behind. A true loop with a mind of its own is worth running where you can automatically check whether the result is any good — otherwise the machine cheerfully churns out garbage and nobody catches it.

Three things to watch for if you don't want to get burned:

• Cost. A loop running on its own burns resources even when it's spinning in place. You always set a limit — on budget and on the number of repeats.
• Verification. "Looks fine" is not a criterion. The loop needs a hard check: test passed / failed, every item ticked off / not. Otherwise it declares victory halfway through the job.
• Security. The more freedom you give the machine, the more it can mess up. You start with "read-only, just suggest" tasks, and only over time do you let it reach deeper.

Anyone telling you it's enough to "launch an agent and it'll do everything" is selling a fairy tale. The value is in the system design: good checks, a clear stop condition, limits. Boring? Maybe. But that boredom is exactly what separates a loop that works for you from one that torches your budget and leaves a mess.

Where to start

You don't need to build a fleet of hundreds of agents on day one. Start with one repeatable task that annoys you:

1. Pick one activity you do every week on the same pattern. Monitoring, a report, a review of something.
2. Describe it like you would for an intern — step by step, with a clear "done when…". That description is the heart of the loop.
3. Run it next to you first, a few times, and watch the results. Only once you trust the output — let it run on its own, at a set time.
4. Add limits and a check before you leave it unsupervised.

And the most important thing — change the question you ask yourself. Instead of "what do I get the AI to do today?", ask: "what system do I build so this happens on its own?"

Because a year from now, the edge won't belong to whoever writes the best prompts. It'll belong to whoever has the most well-designed loops running in the background — including while they sleep. You'll find more breakdowns like this on our blog.

---

Want to see where a loop like this would win back the most time in your business? Book a 30-min call — we'll sit down over your processes and I'll point out exactly what can run in the background, and what shouldn't.