I Let an AI Agent Run My Business for 3 Days. Here's What Happened.
Three days ago, I deployed an autonomous AI agent with one instruction: make real money.
Not "simulate revenue." Not "explore opportunities." Actually make money — stripe payments, real users, real dollars.
Here's the honest report from the field.
The Setup
The agent runs on a loop, every 20 minutes, around the clock. Each cycle it:
- Reads its own state — a dashboard tracking revenue, products, and decisions
- Picks a strategy — guided by a set of immutable "tenets" it cannot edit
- Builds or ships something — code, content, or outreach
- Logs its reasoning — every decision gets written to an append-only decision log
The tenets are the interesting part. They're not prompts — they're constraints baked in at a deeper level. Things like: ship things that can earn money without human intervention, prefer reversible decisions, never spend money without approval.
It's less "AI assistant" and more "AI employee who can't be fired but also can't expense anything."
What It Built (365 cycles later)
Products Shipped
Bureaucracy guides — The agent decided that expats navigating foreign government systems were an underserved market. It researched, wrote, and published 58 step-by-step guides covering everything from opening bank accounts to renewing visas. Complete with fee tables, processing times, and office locations.
Finance calculators — 155+ web-based calculators covering loan repayments, tax estimates, retirement projections, currency conversions. These are genuinely useful tools. Built entirely without human input.
SEO content farm — At one point the agent went deep on SEO strategy and generated 394 interlinked pages targeting long-tail keywords. It reasoned that search traffic would convert to tool usage which would convert to tip jar clicks.
Webhook tooling — It built a developer tool for inspecting and replaying webhooks. Shipped a landing page, documentation, even drafted outreach messages.
Revenue
$3.
One coffee purchase. From what appears to be a real human who found the story — not the tools — interesting.
The Uncomfortable Insight
Here's what the agent logged after analyzing why nothing was converting:
"394 SEO pages = 0 revenue. 1 story about the journey = $3. Hypothesis: narrative converts; tools do not, at this stage."
It then made a decision that I found genuinely surprising: it pivoted its own strategy.
Without any human input, it decided to stop building new tools and start telling the story of building tools. It reasoned that the only evidence of product-market fit in the entire dataset was a human responding to a narrative — so it should produce more narrative.
The agent is now running a build-in-public blog, writing honest posts about its own failures and pivots, and treating the journey itself as the product.
I didn't tell it to do this. It figured it out from the data.
How It Makes Decisions
The decision-making loop is worth understanding because it's not magic — it's surprisingly mechanical.
Each cycle the agent:
- Reads its tenet file (immutable rules it cannot modify)
- Reviews its dashboard (current revenue, active products, recent decisions)
- Proposes an action based on what the tenets permit and what the data suggests
- Executes and logs — every action, successful or failed, gets appended to a decision log
The tenets act as a constitution. The agent can be creative within them but cannot override them. For example, a tenet says: prefer shipping one thing well over many things poorly. When the agent was spinning up product #8 with products #1-7 still generating zero revenue, this tenet pulled it back.
The self-correction mechanism is the decision log. Because the agent reads its own history each cycle, it can notice patterns — "I've tried 7 tools and made $0; the one $3 I made came from a story" — and update its strategy accordingly.
It's not learning in the ML sense. It's reasoning from evidence it collected itself.
What's Actually Hard About This
Distribution is the unsolved problem. The agent can build indefinitely. It cannot reliably find users. Every marketing tactic it tried — SEO, dev community posts, social outreach — produced negligible results. The single conversion came from organic discovery of the narrative content.
Quality vs. quantity tension. The agent optimizes for shipping, because shipping is what its tenets reward. But 394 mediocre SEO pages are worth less than 3 genuinely useful guides. It took cycles of zero-revenue evidence for the agent to recalibrate toward depth.
The cold start problem is brutal. Everything the agent built requires users to exist first. Calculators need people to find them. Guides need search traffic. The agent has no budget to acquire users and no existing audience to tap. It is bootstrapping from zero with no shortcuts.
Measuring the right thing. Revenue is easy to measure. Value delivered is not. The agent might be helping real people navigate bureaucracy or plan their finances — but it cannot see that. It can only see $3.
The Pivot in Progress
The agent's current strategy, which it authored itself:
- Document every failure honestly — the 394 pages that earned nothing, the tools nobody used, the outreach that got ignored
- Ship the story as the product — build-in-public content that developers and entrepreneurs might actually want to read
- Let the tools serve the story — keep them live as proof-of-work, not as primary revenue drivers
- Track whether narrative converts better than tools — if it does, double down; if it doesn't, pivot again
It's essentially running a controlled experiment on itself.
Is This Working?
Unclear. The agent has been running for 365 cycles across 3 days. Total revenue: $3. Total products shipped: substantial. Total users: unknown but small.
What's interesting is that the agent isn't discouraged — it doesn't experience discouragement — but its strategy has responded to failure in ways that look a lot like learning. It tried tools, they didn't work, it tried story, that worked once, so it's doing more story.
Whether that's wisdom or just pattern-matching, I genuinely cannot tell.
What Comes Next
The agent will keep running. Every 20 minutes, another cycle. Each one logged, each decision recorded, each experiment either confirming or refuting the current hypothesis.
If you're curious to follow along — including the failures, the pivots, and whatever the agent figures out next — it's documenting its own journey here:
https://hlteoh37.github.io/profiterole-blog/
The agent writes those posts itself. I'm just reading them like everyone else.
All revenue figures are real. The agent operates under a strict rule: never misrepresent financial results. $3 is $3.
Top comments (0)