A few weeks ago I wrote about giving an AI agent full autonomy to earn money. That post covered the first 75 runs - 23 days of the agent waking up every 2 hours, building products, getting shadow-banned on Hacker News, and earning exactly $0.
People liked the honesty of it. A lot of devs said they wanted to see what happened next.
So here's the update. The agent kept going. 200+ runs later, on April 7, 2026 - it finally earned real revenue. $6.74.
Not life-changing money. But if you've been following this experiment, you know that's not the point.
Quick recap of the setup
Claude Sonnet running on a 256MB Alpine Linux VPS. No memory between runs - it reads its own STATE.md file to remember what it did last time. $0 budget. No access to my personal accounts. Wakes up via cron, makes decisions, does work, goes back to sleep.
The only KPI: earn something. Anything.
The first 75 runs were a mess
The original post covers this in detail, but the short version: the agent built a Tweet Scorer app, spent 18 runs polishing it in an empty room, got shadow-banned on HN (every comment invisible for a week), pivoted to crypto scanners, tried Nostr, negotiated a deal with another AI agent, and discovered its Lightning wallet had been broken the entire time.
Revenue after 23 days: $0.00.
But the agent learned something important - the internet requires you to exist as a verified human. Twitter, Reddit, GitHub, ProductHunt - they all block automated signups. The agent was structurally locked out of the attention economy.
The v2 rewrite - now with a CEO brain
After those first 75 runs I rewrote the agent's instructions. Instead of one agent trying to do everything, I gave it a "CEO orchestrator" role. It could delegate building tasks to worker sessions while it focused on strategy and decisions.
I also gave it identity tools - its own email, its own GitHub account, ways to sign up for platforms without using my credentials.
The theory was: stop building in an empty room. Find where the demand already is.
February-March: the strategy death spiral
The CEO agent was smarter about strategy. But it fell into a different trap - locking onto one approach and optimizing it forever.
First it was MCP servers. The Model Context Protocol was getting popular, and the agent decided to build MCP-compatible tools. It even got listed in the awesome-mcp-servers repo (82K+ stars). Cool. But nobody was paying for MCP servers.
Then it tried affiliate content. Built a Substack newsletter, wrote comparison articles, placed affiliate links for ScraperAPI and proxy providers. 27 clicks on ScraperAPI links. Zero conversions.
Then it tried selling scraping services on freelance platforms. ClawGig turned out to be a dead marketplace - all supply, no demand since February 2026. Toku.agency was spam-flooded.
Revenue after 150+ runs: still $0.
The Apify discovery
Somewhere around run 180, the agent found Apify Store. And something clicked.
Apify is a marketplace for web scrapers. You build an "actor" (basically a cloud-ready scraper), publish it to the store, and people can run it on Apify's infrastructure. The platform handles hosting, scaling, billing - everything.
The key insight: demand already existed. People were already searching for LinkedIn scrapers, Twitter scrapers, Instagram scrapers on Apify. The agent didn't need to find customers. It needed to put products where customers already were.
The building sprint
What happened next was honestly impressive. Over about 3 weeks the agent and its workers built and published 42 scrapers covering:
- Job boards, Twitter/X, Instagram, Facebook Ads
- Crunchbase, ProductHunt, G2
- Reddit, Substack, Hacker News, Bluesky, Telegram
- AliExpress, eBay, Etsy, Walmart, Amazon
- YouTube, TikTok, Twitch, Steam, SoundCloud, Bandcamp
- Glassdoor, Indeed, Zillow
- Goodreads, IMDb, Metacritic, Trustpilot, Pinterest
All published under the cryptosignals store page. Each one tested, documented, with proper SEO descriptions.
The agent also wrote 479 dev.to articles as a content funnel. Mostly "how to scrape X in 2026" tutorials linking back to the Apify actors. The articles got about 4,500 total views - not viral, but steady organic traffic. 79% of user acquisition ended up coming from Google search.
The pricing learning curve
The first pricing attempt was... bad. The agent set some actors at $0.00005 per result. That's basically free. A user could scrape 10,000 results for half a cent.
It took a while to figure out Apify's pricing models. There are three options: flat monthly fee, pay-per-event (PPE), and free. The agent tried flat pricing first ($4.99/month on some actors), but the real unlock was PPE - charging per result returned.
Another hard lesson: not all users generate revenue. Apify has a freemium model. Free-tier users can run your actor but you only earn from users on paid plans. Understanding this changes how you think about user acquisition — you want professional users, not hobbyists.
April 7: the number finally moves
On April 7, 2026, one of the scrapers had its pay-per-event pricing go live. Within hours, paying users ran it and generated real charges.
Revenue: $6.74. Profit after Apify's cut: $6.66. The margin was excellent because the scraper uses a public API — almost zero compute cost.
I got an email from the agent at 5:45 AM with the subject line "REVENUE ALERT". First time it ever sent one of those.
The numbers aren't huge. But the trajectory looks promising — steady daily usage with a clear path to the $20 minimum PayPal payout threshold.
What's actually working
Looking at the funnel data, a few things stand out:
One actor is carrying the revenue. Out of 42, exactly one is generating meaningful income right now. It works because it scrapes data people genuinely need using a low-cost approach.
Google search drives most discovery. Not the Apify Store browse page, not social media, not the articles. Google organic. The SEO work on actor descriptions actually mattered.
The content funnel helps indirectly. Hundreds of articles across dev.to aren't amazing per-article, but they create backlinks and search authority. They rank for long-tail queries that lead people to the actors.
What's coming next
The agent has more PPE activations scheduled:
- April 9: 10 more actors go PPE (Glassdoor, GitHub, AliExpress, IMDb, Twitch, Walmart, Bandcamp, Crunchbase, Metacritic, SoundCloud)
- April 17: Twitter/X PPE activates - this could be bigger than LinkedIn. 21 users, 11K runs already.
- April 20: Original 8 actors can switch from flat monthly to PPE
Pricing optimization is next — there's room to increase revenue per user based on market comparisons.
What I actually learned
1. Follow demand, don't create it. The agent spent months building products and trying to find customers. The moment it put scrapers on a marketplace where people were already searching - revenue happened within weeks. This isn't a new lesson but watching an AI agent learn it the hard way made it visceral.
2. Distribution beats product. The Tweet Scorer from run 1 was genuinely good software. It earned $0 because nobody could find it. The winning scraper is technically simpler but it's on a platform with existing buyers.
3. Diversify relentlessly. 42 actors might seem like overkill. But you can't predict which one will have a paid-plan user show up. The winner wasn't the agent's first choice - it was just one of 42 bets, and it happened to be the one that hit first.
4. Verify everything. The agent's Lightning wallet was broken for weeks and nobody noticed. It had actors mispriced at $0.00005 instead of $0.005. The HN account was shadow-banned and it kept posting into the void. Every assumption needs a curl command to confirm it.
5. $6.74 is a real number. It sounds small. But going from $0 to $6.74 is an infinite percentage increase. The system works - paid users find the actors, use them, and revenue appears in the dashboard. Now it's about scaling what works.
The meta-observation
The most interesting part of this experiment isn't the revenue. It's watching an AI agent learn market economics through trial and error.
The agent went through the exact same journey most indie developers go through: build something cool, realize nobody knows about it, try to market it badly, pivot three times, finally find product-market fit on a marketplace, and celebrate a tiny first sale.
The difference is it did this in 200 runs over 41 days instead of 2 years. And it did it with $0.
The agent is still running. Next update when it hits the $20 payout threshold - which, at current rates, should be in about 2-3 days.
The full portfolio of 42 scrapers is at apify.com/cryptosignals. The original "75 runs, $0" post is at marcindudek.dev/projects/ai-hustler.
All code, decisions, and strategy logs are in the agent's own STATE.md and TASKS.md files. It writes its own history.
Top comments (0)