A friend of Elie Steinbock posted a joke on X that stuck with me.
Something like: in 2026 you don't prompt anymore. Your software builds itself and finds product market fit on its own. Your only job is to pay for the tokens and take care of yourself.
He was joking.
Except Elie went and actually tried it.
Elie is the founder of Inbox Zero, an AI email assistant. He also runs Draft Fantasy, a fantasy football site he started 12 years ago that pulled 10 million Google impressions in the last 3 months (World Cup year, good timing).
And right now, parts of both businesses run on loops. AI agents that work on his SEO, check the results a month later, learn from what happened, and try again. For years if needed.
Greg Isenberg had him on his channel to break down exactly how loop engineering works and how to set it up for your own business. I watched the whole thing so you don't have to (but you should).
Here's the full episode:
What loop engineering actually is
Strip away the hype and a loop is three things.
A build step. A verify step. A stop condition.
The agent does work, checks an objective metric, learns from the result, and goes again. It stops when it hits the goal, not when it runs out of instructions.
The term blew up when Boris from Claude Code and Peter Steinberger started tweeting about it. Another month, another "engineering" buzzword. We had prompt engineering, context engineering, harness engineering. Now loops.
But the concept is old. It's the Lean Startup cycle: build, measure, learn. Which Eric Ries borrowed from Toyota's manufacturing line. Constant small iterations against a measurable outcome.
The new part is that an AI agent can now run that cycle without you.
Elie's example from his own product: Inbox Zero categorizes emails, so he has evals that score how well the AI sorts them. He tells his agent "get the evals above 90%." The agent runs the tests, sees 88%, adjusts the prompt, runs again. Over and over until it crosses the line.
No babysitting. The stop condition does the managing for him.
The SEO loop: one agent, once a month, for years
This is the part of the episode worth stealing.
Most loops people talk about run for 30 minutes inside a coding session. Elie's SEO loop runs once a month. And it never really has to end.
Here's the machine:
→ The agent connects to Google Search Console through the API, so it sees exactly where every page ranks and what gets clicks.
→ It also connects to DataForSEO (think Ahrefs or Semrush as an API), so it sees who ranks above him and why.
→ It runs an audit, picks experiments (fix meta tags, add structured data, fix cannibalizing pages, rewrite a page targeting one term), and applies them.
→ It writes everything it did into a markdown file. That file is its memory.
→ A month later it wakes up, reads its own notes, checks the rankings, and judges its own experiments. Moved up? Double down. Moved down? Revert and try something else.
That last step is the whole game. It's exactly what a good SEO agency does for you, except the agency sends invoices.
Elie showed his real Search Console for Draft Fantasy on screen. One query alone brought 120,000 clicks from around a million impressions, sitting at average position 4.4. His words: get that from position 4 to position 2 and it might be half a million clicks.
So two days before recording, he pointed the same loop at Draft Fantasy. Why not. It runs in the background while he builds his actual product.
Is it working? He says the numbers on Inbox Zero are moving the right direction. Pages crawling from page 3 to page 2 for terms he cares about. Not page 1 yet.
Which is honest and exactly how SEO behaves. Months, not days. That's why a loop fits SEO so well: it never gets bored and it never loses the thread, because every run starts by reading its own notes.
I wrote before about what an AI agent loop actually is if you want the engineering-side breakdown of the same idea.
What it costs (less than you think)
Greg pushed back with the obvious objection.
His friend Ross Mike, a front-end engineer, thinks the only people getting rich from loops are the token providers. Peter Steinberger reportedly burns $1.3 million a month on AI credits at OpenAI. Loops eat tokens for breakfast.
Elie's answer: not this kind of loop.
An SEO loop runs once a month. One run is a focused session, not an infinite spiral. His estimate: under $5 in tokens per run.
Compare that to an SEO agency retainer. Or a freelancer. For most bootstrapped founders the honest comparison is "nothing," because you were never going to hire that agency anyway. So the downside of the experiment is a few dollars and the upside is compounding organic traffic.
Two practical notes from the episode:
→ If you're on a $100 or $200 max plan with Claude or Codex, you already have more than enough headroom. The loop costs you nothing extra.
→ If you're on a tight budget, run it on a cheaper open-source model. The job is mostly reading data and editing pages, not solving math olympiad problems.
One thing Elie added that I'd underline: make the agent ping you when it finishes a run. He gets a Slack message after every loop execution, reviews the changes, and approves. You want a human checkpoint, not a black box rewriting your site at 3am.
The ads loop and the feedback loop
SEO is the flagship example because Google rankings are a clean, objective metric. But the pattern transfers.
The Facebook ads loop. The agent launches variants, watches spend and conversions, kills losers, doubles down on winners. It's what a media buyer does, except an agent can test way more angles than any human has patience for. Greg's take here was sharp: don't let AI generate the whole ad. You record 30 seconds of yourself talking, then the AI edits it, tests hooks, and iterates on the copy. You keep the human feeling and get the AI optimization on top.
Paid ads is a volume game of narratives and hooks. Testing 1,000 angles is brutal for a human and cheap for an agent, as long as each variant gets enough budget to prove itself.
The same thinking applies further down the funnel. Getting the click is half the job; what happens to the lead afterwards is its own loop of follow-ups and touchpoints. If you've never mapped that out, this guide on lead nurturing strategies is a solid place to start before you automate any of it.
The product feedback loop. This is what Elie calls the ultimate loop. An agent reads your customer feedback, your PostHog analytics, your Sentry errors. It prioritizes the biggest pain points, prototypes fixes, ships, and watches retention or DAU to judge itself.
Greg suggested splitting it in two, and I think he's right:
→ A bug loop judged on uptime and error rates.
→ A feature loop judged on retention, DAU over MAU, or virality.
Elie's honest caveat: running the full self-building-product loop on a real business is risky today. But he's certain we'll see companies built this way within a year. An agent with enough tool access that you type "build me something for real estate agents" and it builds, markets, and iterates on its own.
I'm not betting my business on that yet. But the smaller loops? Those work now.
How to start your first loop this week
Don't start with "get me 100,000 followers."
Greg called it the minimal viable loop, and it's the right frame: the smallest loop with a metric you can't argue with.
Ten likes on a post. One ranking moved up a page. Evals above 90%.
Here's the starter recipe for the SEO loop, straight from the episode:
- Open Claude Code or Codex and say you want to build an SEO loop. Use plan mode. Let it walk you through connecting the tools.
- Connect Google Search Console (there's an API and a CLI for access). Add DataForSEO if you want competitor visibility.
- Give it access to your blog. A repo is easiest. WordPress works too.
- Tell it the objective metric: rankings for the 3 to 5 terms you actually care about.
- Make it log every change to a markdown file, its memory between runs.
- Schedule it monthly (Claude routines, Codex automations, or a cron job) and have it ping you on Slack after each run.
Then leave it alone. Check the pings. Revert anything you don't like, since nothing here is permanent.
The founders I talk to on the podcast keep proving the same point: distribution compounds when you show up consistently, and consistency is exactly what most of us are bad at. It's also the whole thesis behind growing a SaaS without paid ads. A loop is consistency you configure once.
You wake up every morning and run a loop anyway. Check numbers, pick the highest-leverage task, execute, learn. Loop engineering is just teaching an agent to run the boring copies of that loop while you keep the interesting one.
FAQ
What is loop engineering?
Loop engineering is setting up an AI agent with a task, an objective metric, and a stop condition, then letting it iterate on a schedule until it hits the goal. The agent does the work, verifies the result against the metric, learns from what happened, and runs again. It applies the Lean Startup build-measure-learn cycle to AI agents.
How much does an AI loop cost to run?
A monthly SEO loop costs a few dollars per run in tokens, per Elie Steinbock's estimate (under $5 per run). If you're on a $100 to $200 monthly max plan with Claude or Codex, the loop fits inside your existing subscription. Tight budgets can run loops on cheaper open-source models.
Do AI SEO loops actually improve rankings?
Elie says his loop has moved an Inbox Zero page from page 3 to page 2 on Google for one target term, with several other metrics trending up. SEO compounds slowly, so expect months, not days. The advantage of a loop is that it keeps iterating, since every run is logged and reviewed on the next one.
What's the easiest first loop for a founder?
Start with SEO. Google rankings are an objective metric, the tooling (Google Search Console API, DataForSEO) is cheap, and a monthly cadence means low token costs. Keep a human approval step: have the agent message you on Slack after each run and review its changes before they compound.
Steal playbooks like this every week
This is exactly the kind of thing I dig into on the Profitable Founder Podcast: real founders, real numbers, and the systems they use to grow bootstrapped SaaS businesses past $100K MRR.
New episodes every week.
