"I'm sweating," I said."I don't understand," the Nest replied.Both lines are real. So is everything that follows.Think. Orchestrate. Act. is the true story of building a voice-driven home assistant from nothing, one real incident at a time - the wrong room that lit up, the meeting room an agent invented out of an unrelated calendar entry, the system that gave advice instead of turning on the fan. Each failure earns a principle: how to describe a tool precisely enough for an AI to use it correctly, when a system should stop and plan before it acts, how to keep an agent's memory useful instead of cluttered, and why a confident answer is never proof of a correct one.The stakes were never abstract. Somewhere in the house, a mother was reaching for a light switch mounted just high enough to be uncomfortable, every morning at 5 AM, rather than ask for help.This isn't a guide to one model or one framework - those will be obsolete before the next edition. It's the judgment underneath them: the same discipline that makes an AI agent trustworthy is the discipline that makes an engineer valuable in a world where agents can write the code themselves.For any engineer who has watched an AI agent do something impressive and wondered what's actually happening underneath - and for anyone about to hand a set of tools to a language model and hope it uses them well.