The first primitive: an agent-agnostic harness
A factory needs agents. This is what an agent is made of, and which part the factory builds.
What an agent is#
An agent is two things: a model and a harness. The harness is everything around the model.
- agent
- modelreads and writes code
- toolsread, edit, run
- contextwhat it knows
- permissionswhat it may do
- loopruns until done
The model is bought. The four highlighted parts are the harness, the part someone has to build.
Reuse, don't rebuild#
Building a good harness is expensive, and good ones already exist: Claude Code, Codex, OpenCode.
- ready–✓
- improves–✓
- my login–✓
- control✓–
So the factory doesn't build one. It reuses whole agents, model and harness included, and gives up some control in exchange. The question is about the process around an agent, not the agent, so that is where the factory's work goes.
The factory's harness#
The factory still has a harness of its own, one level up. It doesn't talk to a model. It talks to agents: it starts one in a repository, gives it the work, and reads what it does.
Any agent can go in that slot. Claude Code is the first one, run from a chat in a desktop app on a local repo.
- factory harness
- claude codetoday
- codexnot yet
- opencodenot yet