More than one agent was too many to keep track of
Each Claude Code session runs in its own Terminal tab. When one finishes or stops to ask me something, nothing tells me. So I mostly ran one session at a time and watched it.
Waiting goes unnoticed
A question from an agent looks the same as a quiet tab. Work stops and I don't know it.
Every tab looks the same
Tabs are named after the folder, so I opened each one to remember what it was doing.
Working or stuck
A long turn of real progress looks the same as one failed command on its tenth try.
Only reachable at the desk
A one-word answer could unblock an hour of work, but only if I was at the laptop.
Each agent's state shown as something a person does
I stop reading status labels after a while, so every state became a pose or a place in the room. I can read it from across my desk.
Where they are says what they're doing
Working sessions sit at a desk. Idle ones wander to the couch or the coffee counter. After half an hour the chair is empty, and a closed session walks out the door.
Only one state is loud
Everything else stays calm so that waiting stands out. When a session needs me, the room dims, a spotlight lands on that desk and the dog goes and sits by whoever has waited longest.
What's inside, without opening it
The monitor shows the kind of step the agent is on. Three failed steps in a row make the desk smoke, which is how I tell stuck from slow. Nameplates show a time estimate based on how long my past turns took.
The conversation opens beside the room
Clicking a desk opens that session's terminal beside the room, so I can read and reply without leaving it. Option-click brings up the real Terminal tab instead. The panel copies what I rely on in Claude Code: the same type and colors, the spinner verbs and timer, and Space to talk.
A light in every session
When the room isn't on screen, a band in Claude Code shows one dot per session: amber when it needs me, green when working, hollow when idle.
Answering an agent from my phone
On a phone the same terminal slides up from the bottom. When an agent asks me to pick from a list, each choice is a row, so most answers are one tap.
A message has to arrive exactly once
Signal drops, and an instruction sent twice can mean the work gets done twice. Replies wait in an outbox and retry, and each carries an id so the laptop never types one twice.
A tap on the wrist
Away from both screens, my Apple Watch taps me. "Needs you" comes after a 20 second wait, "Finished" only for turns of two minutes or more, and neither if I've used the Mac in the last minute.
Changed by living with it
I built the first version in a day and used it while I kept building it.
Side view to isometric
The first room was a flat side view. Isometric gave it depth and space to walk around.
Sleeping to an empty chair
Idle people first fell asleep at their desks. Eleven minutes in, that became an empty chair.
A pane to a band
A pane in Claude Code listing running agents had to be opened to show anything. It went the same day, and the count moved into the band.
Cut the mic button
Phone replies had a mic button until it was clear the keyboard's own dictation already did the job.
It stays one office on purpose, so I passed on a trophy shelf and office upgrades. A queue that started agents for me while I was away was pulled the same day I built it, after Claude Code's safety checks blocked it from launching them.
I run more agents at once now
I compared my Claude Code logs from the 30 days before with my first three and a half days of Agent HQ. Since I started logging replies on October 16, 197 have come from my phone.
It hasn't made me answer faster. The typical wait per session went from 46 seconds to 74, since there are more of them. It's also too early to call a trend, and most of those days were long build days.