I gave an AI agent kubectl, the Argo CD API, my Forgejo token and merge rights on every repository I own, then asked it to bring 178 repos back up to date. Here's what happened, what broke, and what full autonomy without a real verification gate actually costs.
What does it actually look like to work with a personal AI agent day-to-day? Real workflows combining MCP servers, OpenClaw skills, and 4 weeks of production usage patterns. The good, the unexpected, and the genuinely useful.
After building both OpenClaw skills and MCP servers, what did I learn about the right abstraction layer? And where are personal AI agents heading in the next 5 years? A reflection on tooling, standards, and the shift from AI-as-service to AI-as-infrastructure.