Alex Lieberman thinks someone is about to build a world-class brain for enterprises and make a stupid amount of money. He’s right about the money. I think he’s wrong about what you’re buying.
The argument goes: code is easy for AI because context lives in the repo. Knowledge work is hard because context is scattered. So whoever builds the git repo for enterprise context wins.
The first half is right. But the second half buries the actual moat.
We built the repo three months ago. It was the easy part.
Markdown files in git, a folder structure we call ARKS, Claude Code running inside. Grep our entire decision history in 0.3 seconds. Write a decision doc Tuesday, the agent knows about it Wednesday. Full writeup here.
A repo full of context is just a library. It sits there until something reads it with purpose.
A frame that keeps clicking for me.
A company is already a network of routers. The humans are the routers.
Once you see it you can’t unsee it.
Every time our ops lead pulls context from four Slack threads, two Excel documents, a Google Doc, and a thing someone said in a bathroom at offsite 2025. That’s a route. She’s moving context from where it lives to where a decision needs it. We pay her salary for that transcoding. Every company pays someone to do it, whether they notice or not.
Companies don’t fail because the knowledge is missing. They fail because the routes are human. One bus, ten lanes, ten thousand packets, every day, forever. Humans are weirdly good at this. That’s why companies function at all. Also why they seize up when the router quits.
What enterprises are actually buying isn’t the brain, it’s routing capacity. Think of it as inventory versus throughput. A warehouse doesn’t ship itself.
Garry Tan gave us part of the vocabulary: thin harness, fat skills. Model is commoditized. Skill is the moat. A skill is a markdown file that teaches the model how to do one specific thing. The process, not the answer.
What I didn’t appreciate until we’d shipped a few: a fat skill isn’t a workflow, it’s a router.
The skill is a small program that knows the shape of one work function. It reads the right parts of the vault, pulls in the right people, executes the predictable moves, asks a human when the call needs judgment. Same substrate, different routing table, different function out.
Our onboarding router reads your git config, finds your person-note, picks the three meetings you should know about, pulls the relevant decision docs, and runs an interview that enriches your node in the graph. Every step is a routing decision: this context → this person → this moment.
I’ve watched it happen a dozen times now and still get a little flicker of “wait, this actually works.” Nobody tells the agent anything. It just reads the graph.
@contextconor put this sharper than I had it:
The failure mode is not that the agent can’t find information. The failure mode is that it finds too much, can’t tell what’s current, can’t resolve conflicts between sources, and confidently presents a fragment as the whole truth.
That’s exactly the concern. His answer is that agents need something closer to what a great chief of staff has. Context accumulated over months. Knowing what’s current, what’s real, what matters today versus last quarter.
Agents with access to everything don’t have this.
My extension: you don’t need chief-of-staff understanding of the whole company. One work function at a time. A narrow chief of staff.
Conor’s bet is one big brain for the whole company. Mine is many narrow ones, stacked. Same synthesis problem, different architecture. The narrow version is what a fat skill can actually pull off. Which makes it tractable.
Two weeks ago we shipped onboarding. Maria used it. One engineer after her. Each new hire’s person-note enriches the graph, so the next one gets a richer route. Cheap compounding flywheel.
This week we’re building the second one: sales prep.
Context-in: a company name, maybe a contact.
Context-out: who on our side has talked to them, the last three times they surfaced in meetings, which decisions they’ve influenced, whether we’ve committed to anything.
This router is janky. Still hallucinates the occasional contact we never actually met. We’ll fix it.
Third on deck is exec briefing. Brian wants a Monday-morning packet: what shipped, what stalled, what decisions are open, who needs something from whom. Currently a 45-minute manual process. If we hit five minutes and keep it truthful, that’s a week of CTO attention back every quarter.
| Work function | What the router does |
|---|---|
| onboarding | routes context to a new hire |
| sales prep | routes context to a deal |
| exec brief | routes context to a decision |
| hiring loop | routes context to a candidate |
| support | routes context to a ticket |
Same vault. Same markdown files. Different routing tables.
Each skill is a few hundred lines of markdown. None of them are impressive on their own. Together they’re the actual system. That’s the part that gets me excited. Not any one router. The stack.
So: the enterprise brain bet isn’t “whoever owns the repo wins.” The repo is commodity. Markdown in git, we didn’t invent it, and Karpathy didn’t either when he described the same pattern. You could start one today.
The moat is the router stack. Each one a narrow chief of staff for one function.
Every fat skill routes context into work that used to require a human. The vault is the substrate. The routers are the company. Humans get freed up to do the thing they’re actually paid for, which is judgment calls that aren’t in the graph yet.
A year from now, nobody’s asking whether you built an enterprise brain. Everyone will have one. The real question becomes how many work functions your router stack covers, and how good they are. The interesting stuff starts after that, not before.
Things I don’t know.
Some work functions don’t route cleanly. Creative work, hard negotiations, anything where the routing itself requires taste. Those might be routers-in-humans forever. And maybe that’s the right answer. The goal isn’t to route everything, just the parts that shouldn’t have needed a human in the first place.
The understanding gap Conor named is the thing we watch for. Our routers pass it for onboarding because the function is narrow and the signals are structured. They’re going to fail at anything more ambiguous. We don’t know what to do about that yet.
I don’t know the right number of fat skills for a team our size. More than three, less than fifty. We’ll find out.
I don’t know what this looks like at 1,000 people. Different routing problems, different failure modes, different politics. Different post.
But the frame holds at our size. The enterprise brain is a stack, not a product.
Anyway. The first router took us a weekend. The second is taking another. If that math holds, a full stack is months of weekends, not years of building. That’s the part I keep turning over.