how i code

one human, three machines, 8 to 15 agents

i barely open an editor anymore. i answer agents from whatever screen is closest and the system does the rest. this is the whole thing, drawn.

1me. mac, phone, or ipad
3machines. two of them never sleep
8-15agents in parallel, depending on the day
0tmux panes being babysat
the whole thing, one glance
me
boxboxthe pit wall · one inbox
a website. served from the always-on vm below, open on whatever screen i'm holding
macbookcmx serve
boxbox vmcmx serveboxbox lives here
iris's vmcmx serve

cmx serve is the daemon on every machine. the agents live inside it, not in any window, so closing my laptop changes nothing.

every agent stands on: the code itself · pinned tests · the usage log
where i am

me, wherever

i don't watch terminals anymore. when something needs me, it calls me. box box is what the pit wall radios when the driver has to come in right now. i answer from whatever screen is closest.

and on my desk: a dead numpad i brought back, an led under every key. nothing printed on the caps, the light is the label. a permission lands, a key goes hot, i press it once. the wide key is push to talk. i wrote a whole post about it.

the pit wall

boxbox · one inbox for everything blocking any agent, anywhere

the boxbox status bar: 5 calling, 3 driving, $20.45 spent, all three machines up
a permission card off boxbox: cmx wants to run rm -rf node_modules and bun install. approve, always, reject.a review card off boxbox: telemetry v2, boxbox measures itself. plus 2847 minus 12, 9 commits. merge, diff, later.
straight off the actual app, demo fleet. 5 calling plus 3 driving is eight live, if you're counting.
  • it's a chat app. that's the honest category. my agents built it, and keep building it from inside it
  • a website, served from the always-on vm. every machine streams into it over my tailnet, and it follows me to whatever screen i'm holding
  • the real traffic is replies. two thousand a week. about a minute to answer at the desk, eight on the phone. i measured, because of course i did
the fleet

three machines. sessions live in a daemon, not in a terminal window.

my macbookagents work here while i'm in the room, and keep going when i walk away
cmx serve
boxbox vmgoogle cloud, always on. serves the boxbox site, and its agents build boxbox on boxbox
cmx serve
iris's vmiris, my ai teammate. her own email, her own github, her own sim card. long story
cmx serve

cmx is my opencode fork. on every machine it runs as cmx serve, a daemon: starts at boot, never exits, every session lives inside it. terminals, my phone, boxbox, they're all just windows into the same running thing. how it got here: i forked it, then it stopped being a fork.

one session, zoomed

not one ai. an org chart of them.

claude fable 5reads my one message, plans, delegates. the only one that talks back to me
finds how the code works. reads everything, can't write a single file
the heavy stints. real engineering, run in parallelown git checkout
the worker laps. specced edits, cheap and fastasks before deploy
the mule. reads the boring stuff so the expensive ones don't have to
before anything ships: typecheck tests on a real server, no mocks an agent opens the app in a browser and drives it production build

a routing table instead of loyalty. and every writer works in its own checkout, so main stays clean.

the memory

why it compounds

the code itselfthe real memory. a fresh agent doesn't read a summary of what happened, it reads what the last one did, on disk, in actual files
pinned testsevery bug that ever bit me becomes a test against the real thing, no mocks. it doesn't come back
written ruleshonestly the smallest layer: a nudge agents get at the start of a session. it's how i tune the system, not how it remembers
the usage logboxbox measures me using it. one week it counted: 778 prompts, 3,423 marks read, 0 failed verbs. my own usage is the only benchmark that matters

every annoyance becomes a fix, every fix stays fixed. the system gets smarter even on days the models don't.