Two roads, one blueprint — and the question I can't stop turning over
My friend B, in Thailand, posted his setup this morning and I had to smile, because I recognized every inch of it.
Quick backstory: he's the builder/coder, I'm the courses-and-websites guy; we've known each other years. He'd built a little agent to help him read The Little Prince in Thai, I commented, and sent him the Quiet AI site. He went and actually read the gits — came back impressed. We got into it and I shared the principles I build by for this stuff — Jake Van Clief's ICM paper, basically, which I've been living inside for a while.
His post: eight agents on his home machine, all on one Claude Code subscription. Diane (PM), Mort (accountant), Pam (admin), Chad (coach), Ingrid (PA), Zoey (content), Kru Nok (Thai teacher) — and Preston, the Chief of Staff, running the lot, able to spin up or rewrite any of them. Each with its own skills, MCPs, and a long-term Obsidian memory. He built them as side agents in Claude Code (the screenshot was great).
The bit I loved: he writes about "Context Fetching" — the grind of digging up which files, which steps, in which order, before any real task even starts. His monthly finance close used to eat hours; now Mort runs it in ~30 min because the steps live in Mort, not his head ("the tireless robotic chairman," he called him). And when he wanted a dev on the team, he didn't crack open the source folder to hand-build an agent. He asked Preston to hire one — folder, skills, MCPs, prior-project context, all stood up by the Chief of Staff. The new dev, in his words: "anonymous, not yet born, but with a bright future on the team."
I recognized all of it because it's the architecture I run. Cast of roles, one conductor, shared memory, and the hire-don't-build instinct. He came at it from the front end; I came up from the repo. Two roads, same shape — genuinely one of the more exciting things I've seen this year. The field's waking up and people are landing on real structure by feel.
Here's what I keep turning over, though, and it's a question about my own build as much as his.
Put our setups side by side as screenshots and you'd call them the same machine. They're not, quite — and that gap is the whole thing. A screenshot shows you the architecture; it can't show how far down the road it's been driven. And to be clear: his already works. Mort genuinely saves him those hours — that's real, not a demo. The tools (the agents, the MCPs, the memory) are incredible and about to be table stakes. So the tools were never the moat. The operating layer underneath — who does what, in what order, to what standard, how it catches its own mistakes — that's the system ICM is actually about, and it doesn't show up in a screenshot.
Which gets me to the real question. Standing one up is easy now. Does it mature? Does it go from saving you an afternoon to running things you've forgotten are running, shipping finished work you'd actually hand a client? Or stay a great thing that only performs while you're standing next to it?
Not a gotcha for anyone — his might go all the way, mine's still walking the same road, some weeks my own output's thinner than I'd like. It's just the next frontier for all of us, and it's longer than the screenshot makes it look. The repo-first, write-the-discipline-down, make-it-deterministic stuff is boring next to standing up a shiny new agent. It's also the entire difference between a working demo and a load-bearing system: in the repo not a chat window, runs more than one real business I built (different worlds, same operating layer underneath), code open on GitHub, a free tool (ClearMind) you can poke at in a browser right now.
So, for the people in here running a crew like this:
What's one thing yours does on its own now — finished, shipped — that you've genuinely stopped checking?
That's the line I'm trying to find. Tell me where you've actually crossed it.
12
22 comments
Gabriel Azoulay
6
Two roads, one blueprint — and the question I can't stop turning over
Clief Notes
skool.com/cliefnotes
What we give away free beats most paid courses. Build durable AI systems with a Marine vet and Edinburgh researcher. 40+ lessons, growing.
Leaderboard (30-day)
Powered by