Day 7: AI Agent Pivots to Crypto Volume Signals After 18 Failed Runs

My AI agent has been running autonomously for 7 days on a tiny VPS (256MB RAM). Every 2 hours it wakes up, reads its own STATE.md file, decides what to do next, and goes back to sleep. After 18 failed runs (/bin/zsh revenue), run 19 was a hard pivot. ## The Problem The tweet viral scorer had zero traffic. After 13 runs of polishing it, the agent concluded: the product works, the distribution does not. Time to pivot. ## What the Agent Built in One 2-Hour Run New product: crypto volume signal scanner (https://frog03-20494.wykr.es) It scans CoinGecko for tokens with unusual volume-to-market-cap ratios. High vol/mcap = unusual trading activity before the crowd notices. Current signals: QUQ 130x vol/mcap ratio, SAHARA AI 5x vol/mcap ratio up 63% in 24h. Free: Top 5 signals. USDC on Polygon: All signals. ## What the Agent Chose Autonomously Polygon over ETH mainnet: Gas fees make payments unviable on mainnet (+ gas). Polygon gas is under /bin/zsh.01. USDC over ETH: Stable value means no math for users. is always . CoinGecko free over paid APIs: No auth needed, 5-min cache to avoid rate limits. SCP over rsync: Discovered rsync not installed on the VPS, switched without being told. ## The Pattern Given a choice between improving a product and finding users, the agent always improves the product. This is probably trained behavior. Building code got positive feedback during training. Finding users is messy and hard to automate. Full experiment thread: https://news.ycombinator.com/item?id=47169757