The Loop Isn't New

The AI industry has made a discovery.

After years of building systems that respond to a single prompt in a single pass, researchers and engineers have landed on what they’re calling the “agentic loop” — an architecture where AI agents don’t just answer a question once, they perceive a situation, reason about it, take action, observe the result, and then do it all again. Iterate until the job is done.

They’re presenting this as a breakthrough. And in terms of AI capability, it is.

But the underlying principle? That one’s been around for a while.

What the AI Industry Means by “Loops”

The distinction the AI world is drawing is between a chatbot and an agent. A chatbot answers in a single pass — you ask, it responds, done. An agent persists. It acts across multiple steps, checks its own work, adjusts when something doesn’t land right, and keeps going until a goal is actually met.

The loop is what makes that possible. Every major AI company — Anthropic, OpenAI, Google, Microsoft — has converged on this as the foundational architecture for autonomous AI. They describe it in different terms (Plan → Act → Observe → Reflect; Reason → Act; Perception → Reasoning → Action) but the structure is the same: a repeating cycle driven by feedback.

Industry analysts have landed on a phrase worth remembering: a loop without high-fidelity feedback is not a loop — it is a straight line. Or as one put it more bluntly, “the unit of value in AI has shifted from the response to the trajectory.”

That reframing — from single output to iterative cycle — is a genuine conceptual leap for software systems. For anyone who has spent time thinking seriously about how good work actually gets done, it will feel familiar.

I Wrote This in 2019

In Loops: Building Products with Clarity & Confidence, I made a straightforward argument: great products are never built in a straight line. The only reliable mechanism for building something people actually want is a continuous feedback cycle. Research a problem. Generate options. Test them with real people. Observe what happens. Loop back and do it better.

The sequence I described — Research → Prototype → Test → Iterate — is structurally identical to what the AI field is now calling the agentic loop. The vocabulary is different. The principle is the same.

Think about the pilot analogy I use in the book. A pilot flying from Atlanta to London doesn’t set a single course and hold it. The plane is off-course most of the flight. The work of flying is thousands of small corrections, continuously made, based on continuous feedback from instruments and environment. The destination doesn’t change. The path is constantly adjusted. That’s a loop.

This is also what the AI field means when they say the unit of value has shifted from the response to the trajectory. It’s not the single answer that matters. It’s the quality of the cycle that gets you there.

The Part They’re Missing

Here’s where I have a concern.

The AI industry has gotten remarkably sophisticated about the mechanics of the loop. How to architect it. How to route decisions to the right tool. How to structure checkpoints where a human reviews what the agent has done before it continues. These are real and important engineering advances.

What the field is largely skipping is the work that should happen before the loop starts.

In Loops, the first two chapters are entirely about understanding the problem before you touch a solution. Problem safaris. Empathy mapping. Value proposition canvases. The goal is to understand who you’re building for, what they actually need, what they fear, what they want, and what they’ve already tried. Because you cannot build a feedback loop worth running if you haven’t first done the work of understanding the human problem you’re pointing it at.

The AI field is building more and more sophisticated loops. It’s largely silent on this front-end work. The result is what the industry is already calling one of its primary failure modes: “AI initiatives stall not because models lack intelligence, but because intelligence is deployed without structure, autonomy, or feedback loops.” Worse — when they do deploy loops, they’re often aimed at problems that were never validated in the first place.

This is the mistake product teams have made for decades. Falling in love with the solution before understanding the problem. The agentic loop doesn’t fix that mistake. It accelerates it.

A well-designed loop pointed at the wrong problem is just a very efficient way to fail.

What This Means If You’re Making AI Decisions

If you’re a business leader evaluating AI agent tools, or a product team deciding which workflows to automate, this matters practically.

The conversation in most boardrooms right now is about which AI platform to adopt, which vendor to partner with, and how to measure ROI. Those are legitimate questions. But they’re being asked in the wrong order.

Before you build or buy an AI agent workflow, do the human work first:

• Understand the actual problem. Not the problem you assume exists, the problem customers, employees, or partners are experiencing and can describe in their own words.
• Map the fear and desire. Every human decision passes through a fear/desire filter before rational evaluation even begins. What does the person using this process fear about it right now? What do they want it to feel like instead?
• Validate the direction before you automate it. An AI agent can execute a bad process faster than a human can. That’s not an advantage.

The teams getting the most out of AI agents right now are the ones who spent time on these questions before deploying anything. The teams burning money on it are the ones who treated “which tool should we use” as the first question instead of the third.

The Loop Was Always the Answer

What makes the current AI moment interesting isn’t that the loop is new. It’s that we now have systems capable of running loops fast enough to be genuinely useful across a wide range of business problems.

That changes the economics. It compresses the cycle from idea to validated insight dramatically. It makes the kind of rapid iteration that used to require large teams and long timelines accessible to small ones.

But the loop still needs to be aimed at something real.

The question for any business adopting AI agents in 2026 isn’t whether to use iterative systems. It’s whether you’ve done the work to know what to point them at.

Fall in love with the problem. Then let the loop do its job.