All seven roles

Agents

The workers

A brain that has been given a goal and permission to act.

A model on its own can only talk. An agent is a model with a job description: a goal, a set of tools, some rules and a memory. It can read a file, check a calendar, write code or prepare a report, and it keeps going until the job is done.

I run several, because one worker doing everything is as bad an idea in software as it is in an office.

100+

written skills any worker can follow

Who is on the team

  • The personal assistant

    Hermes

    Reachable from my chat apps, day and night. It is an open-source assistant that I run on my own hardware and have extended with nine add-ons of my own.

  • The two senior engineers

    Claude Code and Codex

    They write, review and repair software. For anything important I have one check the other’s work.

  • The junior on site

    OpenCode with a local model

    Works for free on my own machine on small, well-defined tasks. Its work is scored against a written definition of done.

  • The careful one

    A terminal agent

    Asks before changing anything and can be told to plan only. I use it when I want a clear line between suggesting and doing.

  • The specialists

    One checks small business websites. One prepares a daily briefing of security news. One makes a first pass over large folders of photos. One drafts emails: it may propose three versions and is never allowed to press send.

  • The handbook

    100+ skills

    Written instructions for “how we do this here”. Any worker can pick one up, which is how a lesson learned once stays learned.

What managing them taught me

  1. 1

    Workers report success too easily. An overnight run told me six jobs were finished; most were not.

  2. 2

    Give each worker a narrow job and only the tools for it.

  3. 3

    Write the lesson down as an instruction, or the next worker repeats the mistake.

Next

The manager

Decides which brain and which worker gets each job, and how much to spend on it.