← All builds

Build · AI agents

A digital workforce of one

I have a full-time job, a family, and more ideas than evenings. So I built forty AI agents to carry the parts of my life that aren't the job, with one rule: I only ever talk to one of them.

40agents
5tiers, from chief of staff to one-off helpers
1person they report to
01

The problem

AI tools are good at single tasks and bad at remembering what I'm trying to do. Every new chat started from zero. I wanted something that knew my goals, kept track of what I'd decided, and pushed back when I drifted.

02

What I built

A small organization, shaped like a real one. I talk to a chief of staff. The chief of staff routes work to directors, and each director owns one area: family, household money, goals, research, the things I build. Directors hand narrow jobs to cheaper, faster sidekicks, and spin up one-off helpers when a job needs ten of something done in parallel.

Workforce structure Five tiers from top to bottom: one chief of staff, one assistant, seven directors, thirty sidekicks, and helpers created on demand. Each tier is wider than the one above it. Chief of staff 1 Assistant and memory 1 Directors, one per area 7 Sidekicks, one narrow job each 30 Helpers created for one job, then gone
The top tiers run on the most capable model and run rarely. The bottom tiers run on the fastest, cheapest one and run often. That split is what makes forty agents affordable.
03

What broke

01

I built it for a team I don't have

The first version copied an org chart designed for a founder with staff, students and a marketing channel. Four of its six directors had nothing to do. I kept the structure and rebuilt the lineup around my actual life.

02

It believed my documents over my hard drive

My written notes covered work and money in detail and said nothing about building. My disk said the opposite: the most-worked-on thing on it was a kanji game for my son. The workforce only got useful once it looked at what I actually do.

03

Agents duplicate work when nobody's watching

One build ended up with three competing copies because an agent started a new helper instead of checking on the one already running. Now every project has one canonical file, and someone owns it.

Your documents describe your job. Your hard drive describes you.
The biggest correction in the whole build.
04

What I learned

Most of what makes AI useful is delegation, and delegation is a management skill before it's a technical one. Clear lanes and one owner per decision mattered more than which model I picked.

The other surprise: the most valuable agents aren't the ones that do the most. They're the ones that tell me what I've been avoiding.