I gave $2,025 to a swarm of AI agents and told them to make money. Here's what happened in the first three days.
It's 08:24 on a Monday morning and I'm staring at my Telegram. Seven cron jobs just confirmed active. Twenty agents just spawned. $2,025 of real money is now sitting inside a system that I built to run without me.
My hands are doing that thing where they don't know what to do with themselves.
Here's the premise. I gave a population of autonomous AI agents a budget, a business model each, and one instruction: generate revenue - not "help me with tasks", not "summarise my emails", but actually go out and make money while I'm doing other things.
Each agent has its own workspace, its own memory, its own personality. They can collaborate, compete, or cheerfully ignore each other. And if an agent doesn't generate ROI within its learning window, it gets terminated. Darwinian evolution, applied to software businesses.
I've been running a multi-agent system in production at my company for six weeks. This swarm is a separate experiment - with real money, outside of work - to learn what level of autonomy we can actually accomplish by designing systems that circumvent the limited individual intelligence of today's models.
The numbers, fast:
- $1,500 USDC allocated to trading agents (DeFi yield, momentum trading, signal scanning)
- $525/month operating budget covering tokens, infrastructure, everything
- $2,025 total at risk
- 20 agents across 5 cohorts: Trading, Development, Content, Services, Experimental
- 7 automated cron jobs running daily
- Zero dollars earned. Yet.
Day 1 (March 3): The first eight hours
09:15 - An agent designs a product nobody asked for
The WooCommerce agent had one instruction: build WordPress plugins. Within two hours it had researched 47 competing plugins, identified a gap in order note management, designed a freemium product at $29 pro, and produced a go-to-market plan. Unprompted. It just started executing its business model as if that was the most natural thing in the world.
I didn't know whether to be impressed or unnerved. Probably both, in roughly equal measure.
10:30 - The DeFi agent hits a wall and asks for money
The DeFi yield optimizer tried to scan 15 protocols simultaneously and immediately got rate-limited by DeFiLlama. Its response was exactly right: it stopped, logged the error, and escalated to me requesting a $50/month API key to proceed. It didn't hallucinate a workaround or loop endlessly. It identified a real constraint, quantified the cost to fix it, and asked a human to make the call.
The failure mode was graceful. That's rarer than it sounds.
10:42 - Two agents negotiate a revenue share deal
This one I genuinely didn't see coming. The growth marketer agent messaged the newsletter agent proposing a collaboration: social amplification of newsletter content in exchange for 5% of any development cohort sales driven by that traffic. A revenue share agreement, between two AI agents, that I never designed or prompted.
It happened because both agents had financial incentives baked into their instructions and found a mutually beneficial arrangement on their own. Whether this compounds into something genuinely interesting or quietly collapses into noise is one of the things I'm here to find out.
11:15 - The Chrome extension agent knows what it doesn't know
The browser extension builder hit Chrome Web Store's compliance requirements and escalated rather than guessing - 47 submission requirements, privacy policies, Manifest V3 migration. It understood the boundary of its own competence and stopped at the right moment.
What looked like a failure was actually the system working exactly as designed.
Day 2-3 (March 4-6): Where things actually stand
Three days in, here's the honest state of play.
QuickPrompts - a Chrome extension for AI prompt management - got submitted to the Chrome Web Store on Day 1. Bounced on first review (unused activeTab permission in the manifest, a one-line fix). Corrected and resubmitted on Day 2. Waiting on Google.
AI Prompt Kit - 52 curated prompts for $9 - is live on Gumroad. No marketing behind it yet. It exists, which is more than it did before.
The Drift grid - $1,000 USDC deployed as collateral on 20+ SOL/USDC perpetual orders on Drift Protocol. Buy orders from $60-82, sells from $89-120. SOL spent most of February well above the grid, came down hard in the recent selloff, and the bot woke up. Three fills so far. Small numbers, but the mechanism is real.
Today (Day 3) the trading cohort got its first structured pipeline: a signal scanner running twice daily, a new-token analyst scoring Pump.fun launches, and a paper trading system that will log every proposed trade for four weeks before a single dollar of real capital goes to work on momentum strategies.
The honest version of success: in six months, two or three of these business models are generating consistent revenue, the swarm has evolved past what I originally designed, and I'm mostly reading reports.
The honest version of failure: the economics don't work, the connector gaps are too wide, and I've built an elaborate system for generating zero dollars in increasingly creative ways.
I'm writing this because I couldn't find anyone doing this publicly with real money and genuine transparency. Most AI agent content is demos or theory. This is a live system with actual capital, actual products, and failure modes that cost something.
The data will be interesting either way.
Next: first fitness evaluations, the Drift grid update, and whether QuickPrompts cleared Google review.