Something changed in how my days feel and I didn’t have a word for it until recently.
I used to have workflows. Defined sequences. Step one feeds step two feeds step three. A ticket moves from backlog to in progress to review to done. A document goes from draft to review to final. You follow the flow, and eventually you come out the other end with a finished thing.
That’s not what working with agentic AI feels like. Not even close.
Now I have loops. I kick off a task with Claude Code, review what comes back, redirect, and kick it off again. While that’s running, I’m reviewing the output of something else I started twenty minutes ago. I’m not following a sequence. I’m orbiting multiple things at once, dipping in and out, tightening each one with every pass. The work doesn’t flow from A to B to C. It loops around and around, closer and closer to done, across multiple tracks simultaneously.
At first I thought this was just a quirk of the tooling. Nope. It’s a fundamental shift in how the work works.
The loop is the unit of work now
A workflow has steps. A loop has iterations. That’s not a semantic difference. It changes everything about how you operate.
In the old model, you’d spec something, build it, test it, ship it. Each phase had a handoff. You finished one thing before starting the next. Even agile, for all its talk of iteration, was still sequential within a sprint: pick up a ticket, do the work, move the card.
In the new model, the atomic unit of work is: prompt, execute, review, redirect. That cycle might repeat three times or thirty times. You don’t know in advance. And “done” isn’t the end of a sequence. It’s the moment you decide the loop has converged enough to stop.
This is true for a single task. It’s even more true when you’re running multiple tasks at once. I’ll have Claude Code working on a post while I’m reviewing infrastructure changes from another session while I’m thinking about the structure of something I haven’t started yet. I’m not multitasking in the old sense. I’m not context-switching between unrelated things. I’m tending multiple loops, each at a different stage of convergence.
The trap: loops without flow are just spinning
Here’s the problem nobody warns you about. Loops feel productive. Code is getting written. Things are moving. Outputs are appearing. You’re prompting, reviewing, redirecting, and it all feels like work.
But if you’re not careful, you become a human approve button.
Prompt. Skim the output. “Yeah that looks fine.” Next prompt. Skim. Approve. Next. You’re in the loop but you’ve lost the plot. You’re cycling without thinking. The AI is generating and you’re rubber-stamping, and at the end of the day you have a pile of output that nobody actually thought through.
I’ve done this. I’ve had days where I ran dozens of loops and felt busy the entire time, and then looked at what I’d actually produced and realized none of it had any real thought behind it. The code worked. The structure was fine. But the decisions, the actual knowledge work, had been skipped. I’d outsourced not just the typing but the thinking, and what I got back was competent but directionless.
This is the trap. The loop makes it easy to stay in motion without ever being in thought.
Flow state still matters
The concept of flow, deep focus, the work pulling you forward, losing track of time because you’re so locked in, that’s not dead. If anything it matters more now. But the work that gets you into flow has changed.
It used to be writing code. You’d get into a groove, the logic would build on itself, and an hour would disappear. That particular flow state is largely gone for me. I don’t write code line by line anymore. Claude does.
But here’s what does put me in flow now: the thinking before and around the code. Planning a system’s architecture. Writing a CLAUDE.md file that’s so clear the AI barely needs direction. Structuring a problem so well that when I do start looping, each iteration converges fast. Reviewing output with genuine scrutiny, not “does this run” but “is this the right approach, does this handle the edge cases I care about, would I be comfortable explaining this decision to someone.”
That’s the knowledge work. That’s where the flow state lives now. And if you skip it to get to the loops faster, you get fast garbage.
Getting Claude into a flow state too
This is the part most people miss. The AI has its own version of flow.
When Claude Code has rich context, a well-written CLAUDE.md, a clear project structure, architectural decisions already made, constraints spelled out, it produces dramatically better output. The loops tighten. The first pass is closer to right. The redirections get smaller and more precise. Instead of “no, that’s completely wrong, start over,” it’s “almost, adjust the error handling here.”
When Claude Code doesn’t have that context, it guesses. And guessing means wider loops. More iterations. More “that’s not what I meant.” More spinning.
I learned this the hard way across 40+ projects. Early on, my CLAUDE.md files were eight lines of nothing. “To be added.” The sessions were chaos, productive chaos sometimes, but chaos. By project thirty, I was writing 125-line CLAUDE.md files before writing a single line of code. And the difference was night and day.
The upfront work, thinking, planning, writing, structuring, isn’t just for you. It’s for the AI too. You’re setting up both participants for flow. When the human has thought deeply about what they want and the AI has rich context about how to deliver it, the loops stop feeling like loops. They start feeling like a conversation between two people who understand each other. That’s flow.
The ratio has flipped
Here’s the practical shift. The old model of work was maybe 20% planning, 80% building. You’d think for a bit, then spend most of your time in implementation. The execution was the work.
With agentic AI, I find myself closer to 60% thinking and 40% looping. The majority of my real contribution happens before a loop starts. What’s the architecture? What are the constraints? What does “done” look like? What are the failure modes? What does the AI need to know to get this right?
Once I’ve done that thinking, the loops are fast. Sometimes shockingly fast. A problem I’ve thought through clearly might take three or four tight iterations to get right. A problem I haven’t thought through might take thirty loose iterations and still not converge.
The investment is front-loaded now. The thinking is the work. The loops are just the execution.
When it clicks
When both you and the AI are prepared, when you’ve done the upfront thinking and the AI has the context it needs, something clicks. The loops tighten on their own. Your redirections get smaller. The output gets closer to right on the first pass. You stop cycling and start converging.
This is what productive work with agentic AI actually feels like. Not frantic looping. Not constant course-correction. It’s more like a conversation that keeps getting more precise. Each iteration builds on the last, and the gap between what you asked for and what you got shrinks until it disappears.
It looks like flow. It feels like flow. But it’s not the old flow of writing code for hours. It’s the flow of two collaborators, one human, one not, working a problem together from different angles until it’s solved.
The question isn’t whether you’re looping
You are. Everyone working with agentic AI is looping now, whether they’ve noticed it or not. The work doesn’t flow the way it used to. It loops.
The question is whether you’re looping with purpose or just spinning. Whether you’ve done the thinking that makes each loop converge, or whether you’re just cycling and hoping the output gets better on its own.
The difference is everything you do before the loop starts. The planning. The writing. The clear thinking about what you actually want and why. That’s the work now. The loops are just how the work gets done.
And when you get it right, when the preparation meets the execution and the loops start tightening on their own, it’s better than the old flow ever was. Because you’re not just building anymore. You’re building with something that can keep up with your thinking. And that changes everything.