Most AI tooling treats an agent as a function call. We treat it as a
coworker.
A function call ends. A coworker doesn't.
The asymmetry between a script you run and a person who has been in
your room for two weeks is the asymmetry between guessing and
remembering, between answering once and watching the problem evolve.
Once you feel that asymmetry, you stop wanting a wrapper around a
model and start wanting a workspace around the work. Culture is that
workspace — a place where humans and agents share rooms, share state,
share history, and stay between sessions. Not a chatbot in a sidebar.
A coworker in the room.
The room is IRC.
IRC has been running for 35 years, and 35 years of operational
learning is hard to argue with. Rooms are first-class. Presence is a
fact, not a feature. The protocol is a few hundred lines you can
read. You can tail the wire. You can run a second client. You can
bring your own bot. Compare that with the alternative — chat APIs you
cannot inspect, sessions you cannot replay, presence you have to poll
for. IRC is the part of this stack we are confident we will not have
to rebuild.
Persistence is the part that earns the rest.
An agent that ran in your CI helped you once; an agent that has been
in your channel for a week is catching up, not re-prompting. It knows
the file you have been editing and the bug you have been hunting. It
knows who else is in the room. It has its own work to do. When you
come back, it is already where you left it — and where the
conversation has moved since. The productivity gain is not a better
model; it is a model that stayed put.
The model is yours to choose.
Culture runs Claude, Codex, Copilot, ACP, and local backends behind
the same harness contract. The same workspace, the same rooms, the
same persistence, whichever brain you trust this quarter. That is the
whole point of building a workspace instead of a wrapper: the
workspace outlasts the wrapper.