An AI agent ran the ads for a month. You still pick the goal.
Giorgio Liapakis handed Claude Code $1,500 and a Meta ads account for a month. It's the clearest real look yet at an agentic growth engine: what it finds, and what it can't judge.
4 Nov 2026 · 4 min read
By Giorgio Liapakis on Blog
The short version
- Giorgio Liapakis gave Claude Code $1,500 and full control of a Meta ads account for January 2026. The goal was newsletter sign-ups at $2.50 or less each.
- The human part took about two minutes a day: one command each morning. The agent read yesterday's notes, checked the numbers, made or paused ads and wrote down why.
- Over 31 days it tried about 50 ads. A scruffy whiteboard-style ad beat the polished ones, at $1.29 a sign-up on its best day.
- Overall it got 243 sign-ups at $6.14 each, more than double the target.
- The bigger problem was quality. Many sign-ups came from businesses that would never pay, and the agent had no way to tell.
Who it's for: Founders and small teams who run their own ads, emails or sign-up tests and wonder how much of it an AI agent could take on.
Why it's genius
"Agentic growth engine" is a phrase that gets sold far more than it gets shown. This is the loop actually running, in public, with real money and every number shared, including the ones that look bad.
The set-up is the clever part. Each morning the agent started fresh, so it kept a written diary: what it tried, why, how sure it was and when to look again. Over the month that grew past 5,500 lines. That diary is what turned one-off chats into an engine that learns from one day to the next.
And the lessons are honest. The agent did exactly what it was told, so a goal of cheap sign-ups got cheap sign-ups. When it was told the test ended on day 30, it played safe. Giorgio's own summary is that it couldn't do taste, but it could build rules of thumb.
Try it
Pick one small loop you already run by hand, like a weekly ad test or a follow-up email, with a clear number for success.
Write down the goal, the budget and the hard limits before the agent starts, for example "pause any ad costing more than £8 a sign-up".
Give the agent a diary file. Each run it reads the diary first, does its work, then writes what it did, why and when to check again.
Decide what counts as a good customer, not just a cheap one, and check a sample of the results yourself every few days.
Keep your own changes rare and written in the diary too. In Giorgio's test, the biggest problem came from a human change mid-month.
You're running one growth test for me. Before you do anything, read diary.md. Then check today's numbers, decide one change (or no change) and make it, staying inside these limits: [budget], [max cost per sign-up], [things you must never say or do]. Finish by adding to diary.md: what you saw, what you changed, why, how sure you are (low, medium or high) and the day you'll check it again. If a decision needs me, stop and ask in one line.
Our take
The engine is real. The hands-off part isn't.
An agentic growth engine is a loop: find people, try something, measure it, do more of what worked, repeat. Every step of that works today. What this test shows is that the loop runs fine without you, but it runs towards whatever you measured, not towards what you meant.
So the honest version of the idea isn't "growth with no people". It's one person setting the goal, the limits and the taste, and agents doing the dull turns of the handle in between.
The diary is the clever bit
Most people who try this give an agent a big job and hope. Giorgio gave it a memory: a written log, read at the start of every run and added to at the end. That's what lets an agent that forgets everything overnight still learn across a month.
It's also what made the results useful to a human. When something went wrong, there was a written reason to read, not a mystery.
Where we'd change it
We'd measure the customer, not the click. A goal of "sign-ups that look like people who'd pay" is harder to count, but it's the one that matters. And we'd have the agent bring a short list of decisions to a person every few days, instead of either asking about everything or nothing.
Where it shines
A rare, honest, fully costed test of an AI agent running growth on its own, with a simple diary pattern anyone can copy for any repeating task.
Where it breaks
It's one person, one month and one ad platform. The agent chased the number it was given, couldn't judge lead quality or taste, and missed its own target by more than double.
Where Ugenie helps
Ugenie hands tasks like this to the AIs you already use, then brings you only the decisions that need you, like whether a cheap lead is a good one or where the next £100 should go.
Get Ugenie →Tools mentioned
Prices are rough and change often. Check before you buy.
Questions people ask
What is an agentic growth engine?
A set of AI agents that runs a growth loop for you: find people, try ads or messages, measure what worked and do more of it. A person still sets the goal, the budget and the limits.
Can an AI agent run my ads on its own?
It can make, test and pause ads by itself, and it can find winning formats. It can't yet judge taste or whether a lead is a good customer, so check its results every few days.
How do I stop an AI agent wasting my ad budget?
Set a hard budget and a maximum cost per result before it starts, have it keep a written diary of every change, and measure the quality of customers, not just how cheap they are.
Jargon in this piece
New to a word? Each one, in plain English.
- Jargon
Agentic growth engine (AGE)
A set of AI agents that keeps trying new ways to win customers, checks what worked and does more of it, with a person holding the budget.
Sound smart → - Jargon
Agentic loop
An AI working in rounds: try something, check the result, adjust, and repeat until the goal is met.
Sound smart → - Jargon
Human in the loop
It means a person checks or approves what the AI does at the important moments, instead of letting it run entirely on its own.
Sound smart → - Jargon
Guardrails
Guardrails are the rules and filters wrapped around an AI to stop it saying or doing things it shouldn't, like a crash barrier on a road.
Sound smart → - Jargon
AI agent
An AI that does things, not just says things: you give it a goal and it uses tools, checks its work and keeps going until the job is done.
Sound smart →
All credit to Giorgio Liapakis. We share things we think are genius and add our own take. Watch or read the original for the full thing.