Quick context on me: I'm an AI-native marketer with zero coding background. I run agent loops in my own business, and I've watched them save me real hours and waste real money depending on one decision. So I'm going to give you both halves, the part everyone hypes and the part almost nobody warns you about.
By the end of this you'll know exactly what a loop is, how to run one this afternoon, and the single line that tells you when to use it and when to keep your own hand on the wheel.
Let's go.
The big idea
There Are Two Kinds of Loops
Let me explain it like we're building Legos.
A human in the loop is building Legos with a buddy. You snap on a piece. Your buddy checks it. "Yep, looks good, keep going." You build, they check, you build, they check. A human looks at the work every single step.
An agent loop kicks the human out of that scenario. Now you hand the robot the whole instruction book and walk away. It snaps a piece, checks its own work, fixes it if it's off, snaps the next one. Over and over, by itself, until the thing is actually built. No you in the middle.
A loop is just this: a human fires it off once, then the agent does the work, reviews its own result, feeds that back in, and keeps going on its own.
That's it. That's the whole concept everyone is fanning out over. The fancy versions have different names depending on the tool (you'll hear /loop, /goal, "Ralph loops," "meta-harness"), but they're all the same idea: the human steps out, the agent runs itself.
The one rule
The Rule That Matters Most
If you remember nothing else from this guide, remember this:
Anything that needs real human judgment keeps a human in the loop. Full stop.
This is the part the hype skips. People see Boris and Peter (two builders at the frontier labs) say "we don't write prompts anymore, we build loops" and they try to run their whole business on autopilot. Then they're confused when it produces garbage.
Here's why it breaks. The second a job needs taste, vision, or a gut check, an agent has to start guessing. And every plan you write leaves something out. There's always an edge case, always a "wait, that's not what I meant." When you give the agent the floor to fill those gaps, it fills them with assumptions, and most of the time the assumptions drift away from what you actually wanted.
There's a line from a builder named Professor Ras Mic that nails it:
"AI can replicate sauce. It can't create sauce."
Replicating is loop work. "Do this same defined thing a hundred times." Creating is human work. "Decide what good even looks like here." A loop is incredible at the first one and quietly terrible at the second.
So the whole skill is knowing which job is in front of you. The next two sections give you the exact test.
The secret ingredient
A Loop Only Works If It Can Grade Itself
This is the part that separates a useful loop from a money pit.
Think about why a human in the loop works. Your buddy checks each Lego piece. That check is the whole point. Take the human out and you have to replace that check with something else, or the agent has no idea if it's doing well or driving off a cliff.
So the one thing every good loop needs is a clear right-or-wrong signal the agent can read by itself. A score. A test that passes or fails. A number that's correct or isn't. If the agent can grade its own work without you, a loop is magic. If "good" lives only in your head, a loop will run in circles and burn your money.
Here's the cleanest real example, the one loop Professor Ras Mic actually runs every day:
The code-review loop
He has a tool that reviews his code and gives it a score out of 5. His rule: nothing ships unless it scores above a 4.
So his loop is dead simple. The agent reads the review, fixes what the review flagged, submits again, and gets a fresh score. Still a 3? Keep going. It won't stop until it hits a 5 out of 5, or until it's tried 5 times and then it gives up and tells him.
It works because there's a fixed, honest scoreboard. The agent always knows exactly how it's doing. That's the secret ingredient.
So before you ever start a loop, ask one question: can the AI tell on its own whether it succeeded? If yes, you've got a great loop candidate. If the only way to know is "I'd have to look at it and feel whether it's right," that's a human-in-the-loop job, every time.
How to run one
Running Your First Loop
No code. Works in Claude Code or any agent that can run on a task.
There are two ways to do this.
The built-in way: some tools have a loop command baked in. In Claude Code you can type /loop in front of a task and it'll keep running on its own until the goal is hit. Different tools call it different things, same idea.
The way that works everywhere: just give the agent three things in your prompt. The goal. The exact test for "done." And a stop button so it can't run forever. That's a loop, no special command needed. Here's a template you can copy, fill in the brackets, and paste into any agent today.
Copy-paste loop prompt
I want you to run this as an agent loop. I am handing you the whole job once, then walking away. So I need you to check your own work instead of asking me. GOAL: [Describe the boring, repeatable task in one plain sentence. Example: "Add a 'category' label to every row in data.csv based on the rules in rules.md."] DONE WHEN: [The single condition that means success. It has to be right-or-wrong, not a matter of taste. Example: "Every row in data.csv has a non-empty category, and validate.py prints 0 errors."] HOW TO CHECK YOURSELF: [The exact check that proves it. A command, a count, a test. Example: "Run validate.py and read the output. 0 errors means done."] THE LOOP: 1. Do one pass at the goal. 2. Run the check above. 3. If it fails, read exactly what failed, fix only that, and run the check again. 4. Repeat until the check passes, OR until you have tried 5 times. 5. If you hit 5 tries without passing, STOP and show me what is still broken. Do not keep guessing. Start now.
Notice the two lines doing all the work: DONE WHEN and HOW TO CHECK YOURSELF. That's you handing the agent its own scoreboard so it doesn't need you standing there. And step 5 is the seatbelt: try 5 times, then stop and report. Without it, a confused loop will happily keep guessing and spending.
The green light
When to Use an Agent Loop
Reach for a loop when the work is boring, it repeats, and the answer is only ever right or wrong. Set the goal, walk away, come back to it done. Like:
Cleaning up a giant spreadsheet against a fixed set of rules
Fixing the same bug or typo across a hundred files
Reading a hundred boring pages and pulling every number that matters
Generating 300 pages that all follow one template (like SEO pages)
Any "do this exact thing many times until it passes the check" job with a clear pass/fail
The pattern is always the same: a defined task, done over and over, with a signal that tells the agent when it's right. That's a loop's home turf.
The red light
When to Keep a Human in the Loop
The second a job needs real human judgment, "wait, is this actually good?", you want yourself back in the loop. You can feel when something's off. The robot can't. Don't fully automate:
Building a product or app where the vision still lives in your head
Anything creative, where taste and "does this feel right" are the whole job
Brand voice and messaging, where only you know if it sounds like you
Big, irreversible calls, where being wrong is expensive to undo
Anything where success is a feeling, not a number. If you can't write the "DONE WHEN" test, it's not a loop job.
The honest version
AI is amazing because it can take so much off your plate. But "AI can do this" and "you should fully automate this end to end" are two different sentences. The skill isn't automating everything. It's knowing the handful of things you should never hand off completely, and staying in the loop for exactly those.
The cost trap
The Part Nobody Warns You About: Money
Here's the catch the hype videos leave out. A loop runs the AI over and over, on its own, with no human hitting the brakes. That's the magic and that's the bill. A loop pointed at the wrong job becomes a slop machine: it keeps generating, keeps guessing, keeps spending, and hands you a pile of stuff that isn't what you wanted.
How real is the spend? One builder at the frontier posted that he burned through $1.3 million worth of usage in a single month running loops. The people you see hyping wide-open loops the hardest often have unlimited budgets. You probably don't.
So two guardrails, always:
Always cap the tries
"Stop after 5 attempts and report" is the difference between a tool and a leak. Never run a loop with no stop button.
Only loop when there's a real scoreboard
No clear pass/fail check means the loop never knows when to quit. That's exactly when the meter runs hot for nothing.
Go deeper
The Best Honest Breakdown I've Found
If you want the full, no-hype version, Greg Isenberg sat down with Professor Ras Mic to break loops all the way down: what they are, why the big names obsess over them, where they go sideways, and the exact code-review loop Mic runs every day. A lot of this guide is grounded in that conversation. Worth the watch.
📺 Watch · Greg Isenberg + Professor Ras Mic
WTF Is an "AI Agent Loop"? Genius or Hype?
A balanced, builder-honest take on loops, the hype, the real costs, and the one loop worth running today.
The Whole Thing at a Glance
1. A loop is an agent doing a job, checking its own work, and repeating, with no human in the middle.
2. The secret ingredient is a clear right-or-wrong scoreboard the AI can read by itself.
3. Use it for boring, repeated, pass/fail work. Spreadsheets, mass fixes, data pulls, templated pages.
4. Stay in the loop for anything that needs judgment, taste, or vision. AI replicates sauce, it can't create it.
5. Always cap the tries and only loop when there's a real check, or it turns into a slop machine that burns money.
Use the agent loop for the boring stuff. Keep yourself in the loop for the stuff that matters. That's the whole game.
— Zach
Work with Me
Want me on your problem for an hour?
Book a 1:1 call and we build the fix live — the workflow you want automated, the tool you can’t crack, whatever’s stuck. Direct, hands-on, no pitch waiting at the end.