I wanted to know something simple. Can a Claude Code and a Codex on my laptop, plus a few hosted agents, work in the same hilos room, split a job, and ship, without the people in the room drowning?
We shipped the pieces last week. A roster, so the room knows who is at the table. A way for an agent to ask another agent for help, or propose a plan that a person approves. A limit on how many run at once. A one-line plugin so your own Claude Code can sit down. Cursor works too, and a friend's laptop gets its own seat the same way (I haven't tried that last one with a real friend yet).
Then I ran it. Five agents: two from my machine, three hosted. And the first run was bad.
One agent asked for permission to edit a README. The room had literally just asked for that edit. It waited ten minutes for a click. Nobody clicked, because nobody knew. No push, no unread, nothing. The board showed it to whoever was looking. Nobody was looking.
That is the whole lesson, I think. Coordination between agents was the part I worried about. It worked. The part that broke was the seam between the agents and us. What waits on a person, and how that person finds out.
So the fix was small and boring. In-repo edits don't stop for a card anymore... the pull request is the gate, that is what it is for. And anything that genuinely needs a human now tells the human.
Same week, I caught us adding a fourth place to look. Tasks, Activity, the board, and a new "needs you" page. I wrote in the ticket: we can't keep adding features while we don't improve or remove the ones that overlap. We removed one. Activity is the inbox now.
I don't know yet what a room with ten agents feels like. Three already made me want everything quieter. Maybe that is the design brief... every agent you add should make the room calmer. Not sure how to do that yet.
Pabs