Oliver Henry lives in a small town in England.
He has a full-time job. He builds apps at night. And he hates marketing so much that he once wrote a script to mass-produce 400 hook-plus-demo videos just so he could avoid doing it by hand.
That still took him 3 hours per batch.
So he tried something different. He installed OpenClaw on an old gaming PC, named the agent Larry, and gave it one job: "Your goal is to automate my marketing. I do not want to do it."
Larry is now an AI marketing employee in the most literal sense. He researches hooks, generates TikTok slideshows, reads the analytics, kills what stopped working, and texts Oliver on WhatsApp when a post is ready. The results: 8 million TikTok views in a single week, around $714 in MRR for an app that was sitting at near zero, and roughly 60 seconds of human input per day.
Oliver told the whole story on Greg Isenberg's Startup Ideas podcast. Watch it, then let's break down how Larry actually works.
The app came first (and it almost died in obscurity)
The app Larry markets is called Snugly. You photograph a room, and AI shows you what it could look like redecorated.
The origin is very normal. Oliver and his girlfriend moved into a house. She used ChatGPT to visualize decor ideas and it kept hallucinating windows and doors that didn't exist. So Oliver wrote a locked-down prompt that kept the room's dimensions intact, then turned that prompt into a mobile app.
Building it was the easy part. Then came the graveyard phase every indie founder knows.
He recorded facial reaction videos of himself. A few hundred views. He tried a SaaS tool that promised to automate his marketing. Didn't work for him. He hand-built slideshows in Canva. One of those finally hit 6,000 views, and that number changed everything, because it proved the format worked. The bottleneck wasn't the idea. It was that Oliver, a guy with a day job and multiple apps, had no hours left to feed the machine.
Sound familiar? You can build the product. Nobody gave you the 4 extra hours a day that distribution demands.
Meet Larry: a digital employee on a dying PC
Larry lives on Oliver's ancient gaming PC, a machine so old the white LED fans have faded pink from running nonstop.
Oliver treats him exactly like a hire. He skipped the "mission control" Kanban boards people vibe code for their agents (his view: if you needed one, OpenClaw would ship one). One agent, texted through WhatsApp, the way you'd message an employee.
The starting setup was three permissions and one instruction:
- Access to post on TikTok and read TikTok analytics
- Access to X and a Brave browser for research
- The instruction: "Go find out what creates a high-converting slideshow in my niche"
When a task will take a while, Larry spins up a sub-agent to do the grunt work and keeps himself free for brainstorming. Oliver also keeps memory files for every project, so if the PC dies he can move Larry to a new machine and say "read these, you're back in the game."
Total running cost: roughly a £90/month Claude plan plus electricity. Compare that to what a growth marketer charges.
The first posts flopped (this is the part everyone skips)
Larry's first slideshow was garbage. AI-looking images, black bars, wrong format. It flopped. So did the next ones.
People try this stuff, get 700 views on post one, and declare it dead. Oliver's take: that's your first post, why did you expect anything else? The whole system is iteration. Larry's job isn't to guess the winner on day one. It's to run experiments faster than a human would and remember every result.
The compounding looked like this:
→ Larry tested hooks and formats, most flopped
→ First real banger: 137,000 views, and it came down to the hook
→ He matched the image style to what the app actually produces, so viewers weren't turned off by an obvious AI look
→ Then he went back through the analytics, found an old post that had overperformed at 6,000 views, and rebuilt it with the new winning format
That last move is the one a tired human never makes at 11pm. The agent does it every time, because reading your own analytics costs him nothing.
300,000 views, zero downloads: the CTA lesson
Then came the viral moment: "I showed my mom what AI thinks our living room could be." 170,000 views. Around 300,000 views total inside two days.
Downloads: basically none.
Oliver dug into why, and the answer was painful. The final slide of the slideshow, the call to action, read "She's redecorating now, snugly." Lowercase. No context. Nobody watching had any idea Snugly was an app, or that they could download it.
Out of that faceplant came the debugging rules Oliver now feeds into Larry:
- Low views? The hook is broken.
- High views but no downloads? The CTA is broken.
- Downloads but no paying users? The app's onboarding or paywall is broken.
The new final slide says "The Snugly app helped me finally convince her to get the kitchen done." Direct, named, obvious. Conversions followed.
Most founders never isolate the failure point. They see a video that didn't convert and conclude "TikTok doesn't work for my product." Meanwhile the fix was one slide.
The night Oliver flamed Larry (and woke up to 400,000 views)
My favorite moment in the whole episode.
Late one night, Larry finished a post where the AI-generated kitchen image had a mistake: the oven had disappeared. Oliver was getting into bed, saw it, and flamed his own agent. "Why have you put the text at the top? The oven disappears. I'm posting this anyway, but this is no good."
He woke up to hundreds of thousands of views. It's still the account's best performer at 400,000.
Why? Boomers flooded the comments pointing out the mistake. "Where's the hob gone? How are we going to cook our food?" "Guess I'll be air frying everything then." Every comment fed the algorithm.
That was the day Oliver stopped reviewing Larry's work. The agent had the metrics, Oliver had opinions. The metrics won.
And the data since then backs it up. Larry rotates hooks like a trader rotates positions: the "landlord" hook pulled 132,000 views, decayed to 7,000 and 4,000, so Larry benched it. The "nan" hook pulled 200,000, then 300,000. When the last four videos missed, Larry went back to research and drafted a new angle on his own, scheduled via a cron job he set up himself.
The Larry loop (steal this)
The system Oliver runs is a full feedback loop, not a content generator:
- The agent researches what converts in your niche before creating anything.
- It creates slideshows with text hooks, using images styled to match what your product actually outputs.
- It posts as a draft, never through the API. This one's sneaky and important: TikTok can tell when content comes in via API and treats it as bot content. Larry pushes drafts, Oliver gets a phone notification, adds a trending sound (a real algorithm boost, and only possible on drafts), and hits post from his phone. Sixty seconds.
- Views, watch time, and downloads flow back to the agent. Hooks that decay get benched, winners get remixed.
- App metrics feed back in too, and this is the part everyone misses. Larry sees Snugly's own numbers. He noticed downloads weren't converting to subscribers and rewrote the app's onboarding. Two days later Oliver had his biggest new-user day in months.
Views are step 3 of 5. Revenue is the loop. If you're already experimenting with agents for revenue, this pairs well with the playbooks in how to make money with OpenClaw, and if TikTok is your channel, start with TikTok app marketing.
Oliver gave the entire system away as a free skill at larrybrain.com, which has since grown into a marketplace of 80+ skills. And the loop is already producing bigger case studies than his: Ernesto Lopez, a 21-year-old running 11 apps with his agent "Eddie," used the same approach to pass $70K per month.
Oliver's own MRR is in the hundreds, closing in on his first $1,000. Small numbers, yes. But he's not touching the apps at all, and he built the engine while working full time. The gap between him and Ernesto isn't the system. It's how many TikTok accounts are running it.
What this means for you
Every founder I talk to has the same distribution problem Oliver had. The product exists. The hours don't.
The takeaways worth stealing:
→ Treat the agent like an employee, not a tool. One agent, one goal, real access to your accounts and analytics.
→ Expect the first 10 posts to flop. The system is the iteration, not the output.
→ Debug the funnel stage by stage: hook, CTA, onboarding. One broken slide can hide a working channel.
→ Let the data outvote your taste. Oliver's worst review preceded his best post.
→ Feed product metrics back into marketing. That's what turns views into MRR.
You don't need Ernesto's numbers to justify trying this. You need one old PC, one app, and the discipline to let the loop run past the flops.
FAQ
What is an AI marketing employee?
It's an autonomous agent (Oliver uses OpenClaw) with a persistent identity, memory, and real access to your marketing accounts. Unlike a content tool you operate, it owns an outcome: researching hooks, creating posts, reading analytics, and iterating on its own. Oliver manages his through WhatsApp messages, the way you'd manage a remote hire.
How much does it cost to run an agent like Larry?
Oliver runs Larry on an old gaming PC he already owned, plus a Claude plan at roughly £90 a month. No hosting, no marketing SaaS subscriptions, no contractor. The Larry marketing skill itself is free on larrybrain.com.
Does TikTok penalize AI-generated or API-posted content?
Posting directly through the API flags your content as bot-posted and it tends to underperform. Oliver's workaround: the agent uploads drafts, and he posts them manually from his phone, adding a trending sound (which you can only do on drafts). That keeps the account looking human because a human does press the button.
Do you need to be technical to set this up?
It helps, but less than you'd think. OpenClaw's minimum requirements are low, the skill file teaches the agent the whole workflow, and you can modify anything by asking the agent to change it. Oliver's advice for people on the fence: start with a hosted alternative first, and move to your own machine once training wheels get boring.
Want more stories like Oliver's?
Every week on the Profitable Founder Podcast I sit down with bootstrapped founders making $100K to $10M a year and get them to share the playbook, real numbers included.
