Six months ago I stopped playing with AI agents and started living with one. It runs 24/7, has persistent memory, sub-agents for dedicated roles, integrations with Gmail, Calendar, docs, and 15+ capabilities. Most of the time it works. Sometimes it breaks in ways I didn't expect. This talk covers the architecture that emerged, three concrete failures (including one where the agent replied on my behalf about a sensitive matter), the four principles that came out of the mistakes, and one rule that guards against most problems. Not a "how to build an agent" talk. A "what happens when you actually live with one" talk.
Software Engineer, AI Agents Specialist