Starting Cadence

“I woke to neither an alarm nor my mom’s shouting, but to the cool morning air and the pale light of a sun not yet risen. Unzipping the tent, I watched sunlight break through the clouds above the horizon, scattering gold across the restless sea. Months later, I awoke to the screeching from “The Immigrants”. The room was still dark, yet the clock read 8:15. Half-awake, hair half-dried, bag half-open, I ran to school, only to arrive late.”

Same body, same brain, so what was the difference? That question is what kicked off this whole thing.

It started with circadian rhythm. How I wake up seems to matter a lot more than I ever gave it credit for, and getting ripped out of sleep by a screeching alarm is a pretty good way to start the day in fight-or-flight mode before I’ve even opened my eyes. (The 8:15 sprint was a live demo.)

So thought one: what if I woke up to a song instead? Not something that kicks the door in, but something that eases me into the day and makes mornings a bit more pleasant.

Thought two: a song by itself is kind of dumb. It doesn’t know when I actually need to be up, or what my day looks like. The song and the schedule should be connected.

Thought three, which took me embarrassingly long to land on: if those two should be connected, why am I not connecting them with an AI agent?

So I went down the agent rabbit hole. I played around with Hermes agent and a handful of other kinds of agents. They’re cool, but they come with a lot: tools, plugins, layers on top of layers, and most of it was stuff I’d never touch for this.

Then I asked myself what the actual job is. Speaker as input, the agent decides, something comes out. That’s the whole thing.

So I built exactly that and nothing else. A simple tool-using AI agent where I stripped out everything in the middle, all the extra tools and plugins, and kept the main harness as small as I could get it.

The catch with going minimal is getting stuck there, so I designed the architecture to be easy to expand. When I want a new tool, it slots in. When I want a plugin, same deal. The core stays simple and the edges can grow.

The last piece was integrating Jev for the decision step. It makes the agent’s decisions faster and more accurate at the same time, which is the part I’m happiest with, because usually you have to pick one.

That’s where Cadence is right now: a small, simple agent with room to grow, built to expand on to the speaker and lights. Just a simple beat and nothing more.