A lot of the messages I get ask the same question: how do you actually build agents?
I built twenty-two of them. Or tried to.
The vision was clean. Multiple agents, each with a job. A hierarchy above them to coordinate. A chief of staff at the top. Autonomous. Orchestrated. Running while I slept.
I had it all mapped out.
What I actually built was infrastructure. State files. Message queues. Classification layers. Trust ladders. Approval frameworks. For months, that was the work.
The hierarchy collapsed into a single interface. The twenty-two agents became thirty-eight jobs, none of them consolidated. The outreach agent had eight hundred and eighty-one contacts queued and ready. It never sent one. The button never got pushed.
At some point I wrote a note to myself: return comes from using the system, not extending it.
The most honest line I wrote about the whole thing:
The system was built to reduce my decision load. The biggest thing it produced was an approval interface for me to make more decisions.
Here is what I learned from all of it.
The wrong database. The wrong agent structure. The wrong problems to solve. None of it was the technology's fault.
I did not understand AI well enough to use it properly. I understood enough to start building. Not enough to build the right thing.
I needed to go back to school. Not because the capability was not there. Because I had not done the work to understand what the capability actually was, and what it was not.
That is where the next part of this goes.