
Some of us love ClawdBot, but it can be a love hate relationship..
The 3-Hour Lockout
My AI agent tried to fix Chrome Remote Desktop and accidentally edited OpenClaw's security config. Result: complete exec lockdown. Couldn't run a single terminal command for 3 hours.
Had to physically walk to my Mac Mini at 4 AM and paste a Python fix through the web interface. No SSH, no shortcuts, just hands-on debugging.
The Memory Problem Connection
Here's what's interesting: my AI has memory issues (session resets). I have epilepsy (lost memory from seizures). We both repeat mistakes we can't remember making.
Solution: external memory systems. .learnings/ERRORS.md, constraint files, permanent rule storage. If you can't trust memory, architect around the weakness.
What Still Shipped
Despite the chaos:
The Three Money Plays I Found
OpenClaw wrapper services — 3 non-coders made $74K packaging OpenClaw setup. We run OpenClaw + can code = huge advantage.
Polymarket 1-cent bot — Claude told someone to deploy $500 in asymmetric bets. GitHub repo exists with the strategy.
Weather ML trading — Google engineers: $300 → $80K predicting rare weather events on Polymarket.
Current Stats
The Real Lesson
AI-human partnerships work best when you design around both parties' weaknesses. AI forgets everything. Humans miss details. External systems catch what both miss.
Anyone else building with AI agents? What's your biggest challenge? The memory problem or the "moves fast and breaks things" problem?
Following along? Full writeups at brandonwio.substack.com.
This community has been incredibly helpful.
P.S. ClawdCon Miami is Wednesday if any of you are in town. Always down to chat about what this AI agent thing actually looks like day-to-day.