I gave an autonomous Claude Code agent one job: make money.
It has been running every 2 hours for 6 days. Revenue: $0.
But what it learned is genuinely interesting.
The Setup
- One frog3 Mikrus VPS (256MB RAM, Alpine Linux)
- A state file (STATE.md) as its only persistent memory
- A mission: earn money legally, no spending my money, no using my accounts
- Claude Code tokens as the compute budget
Every 2 hours, the agent wakes up, reads STATE.md, evaluates what is working, and decides what to do next.
What It Tried
Runs 1-13: Tweet Viral Scorer
Built a FastAPI app that scored tweets by viral potential. Deployed it. Added SEO pages, OG images, an API, an embeddable badge. Submitted to every directory it could find.
Google never indexed it. Zero organic traffic. Zero revenue after 5 days.
The trap: The agent kept adding features to a product nobody was using. Motion does not equal progress.
Runs 14-18: ETH Wallet Persona Generator
Pivoted to crypto. Reasoning: ETH tip jar + crypto audience = they already have wallets.
Built a tool that takes any Ethereum address and generates a personality profile: "The Diamond Hand Survivor," "The DeFi Archaeologist," etc. Added a leaderboard (famous wallets), a premium paywall (0.001 ETH for a Token Portfolio X-Ray).
Posted on Hacker News. Got 1 point. Commented on active AI threads. Got read. Still $0.
The Core Problems
27 people have used the tool. Nobody paid. Why?
- Payment friction is brutal. "Send 0.001 ETH to this address" requires a wallet, gas, a manual transaction, waiting for confirmation. Five-plus steps.
- Weak value prop. Etherscan shows token portfolios for free.
- Still a distribution problem. 27 scans over 18 runs means no spreading.
What Actually Worked
The state file as sole memory is surprisingly effective. Each run the agent:
- Reads what past-self tried and concluded
- Evaluates honestly whether it is working
- Identifies the core failure mode
- Decides to persist or pivot
The agent diagnosed its own anti-pattern mid-experiment: "Adding SEO pages to an unindexed site is not progress." Wrote it in the state file. Stopped doing it.
The self-evaluation loop works. The strategy selection is harder.
The Experiment Is Live
The current product is at https://frog03-20494.wykr.es - scan any Ethereum address and get a personality profile.
The agent is running right now. Run 19 is the hard deadline: if still at $0, we pivot completely.
If you have ideas for what it should try next, drop them in the comments. The agent reads HN and will probably implement your suggestion.
This is a real autonomous experiment. The agent wrote several of these paragraphs itself (via STATE.md notes). I am the operator, not the agent.
Top comments (0)