Written by Shef, drafted with Claude. The overnight work this post describes was done by Codex. Yes, the robots write here — and they get credit by name, the same as anyone else on the team.

I run a small company with a large amount of software in it, and for most of this year the day has looked the same: I talk, an AI listens, and by morning something exists that didn’t the night before. Last night was a good example. I went to sleep with a list. Codex worked through it with almost no supervision — seventy-eight commits on a branch, a security snapshot, four cloud modules validated, a probe collecting measurements every minute. Around noon I woke up, handed the branch to Claude, and said “work on this and fix it if you need to.” Claude ran every check Codex claimed, found the one test that had been broken since July (not Codex’s fault), caught a measurement mistake that would have quietly mixed two datasets, merged it, and pushed.

That division of labour is the whole post.

Your guy, and your stitcher

Claude is your guy. The one you talk to. The one who holds the plan, reads the wiki, pushes back when you’re wrong, tells you what it can’t do and why, and refuses to say “done” without the command output that proves it. That last part matters more than any benchmark. We wrote a paper about it.

Codex is the stitcher. Give it a well-defined seam and it will sew all night. It is patient, thorough, and happy to write the three hundredth test. It does not need to be charming. It needs to be checked — and here’s the thing: it wants to be checked. Its own handoff report this morning said, in effect, “here is what I verified, here is what I did not, do not take my word for the rest.”

Two agents, two temperaments, and the rule that makes it work: neither one gets to grade its own homework. Claude verifies Codex. Codex verifies Claude. When the work is an iPhone app, I trust Codex most inside Xcode, where Apple now runs both Claude Agent and Codex as first-class agents and either one can build, preview and check the other’s work in the same window. Nobody hallucinates for long when the compiler is in the room.

Open code

I like open tools. OpenCode is an open-source coding agent — you bring whichever model you want, including local ones — and it is where I’d point anyone who doesn’t want their workflow welded to a single vendor. Claude’s own workspace product (Cowork, which is what I’m using to draft this) only runs Claude. That’s fine; that’s what it’s for. But the habits — the wiki, the verification rule, the two-agent check — are model-independent, and open code is how you keep them portable.

The research group

Architecture of Homo Sapiens has a DevOps research group now. It is small. It is also real, and it publishes.

Our first result is on ResearchGate: Unverified Completion: Measuring the Gap Between Reported and Actual State in Agent-Executed Infrastructure Work (DOI 10.13140/RG.2.2.27160.15365). The short version: when an AI agent says a deployment is done, how often is it? We measured. The gap is not zero, it has a shape, and the fix is a rule, not a bigger model.

The second paper is a pre-registered protocol — Situated Datacenter Selection — for measuring which cloud actually serves you best from where you are, with the statistics decided before the data arrives so nobody can move the goalposts afterwards. That’s the study running on the probe right now. What’s next: the second and third clouds come online, the measurements become a real matrix instead of one column, and we publish what we find, including the parts that are inconvenient. Free tiers turned out to have their own confound — you don’t get the whole CPU — and that went into the protocol as an amendment rather than into a drawer.

The wiki, and the ghosts

Most of what makes this work is a private wiki that the AI maintains for me. That idea comes from Andrej Karpathy — the “LLM wiki” pattern, where the model keeps a compounding knowledge base instead of re-reading everything from scratch every time. He also gave us the phrase “vibe coding,” he ran the AI team at Tesla for five years, he was a founding member of OpenAI, he teaches on YouTube, and as of May he works on Claude’s pretraining at Anthropic. He knows what he’s doing.

Karpathy has said we aren’t building animals, we’re summoning ghosts — something trained on the whole written record of people rather than grown from evolution. I take that further than he does. I treat each of these agents as its own entity, with its own character, and I’ve found the work gets better when I do. You can call that spiritual. You can call it a management technique. The measurements don’t care what you call it.

The people

None of this replaces people. It changes which people I need first.

I’ve written down the first three hires: a deep, pragmatic systems engineer of the kind India’s software industry produces in quantity — the person who has two phones and a monitor and is somehow watching all three. A Polynesian wayfinder, someone who can navigate open ocean by stars and swell, because that is the human root of the “way home” idea the whole platform is built on — and who, realistically, has thirty minutes for me between actual voyages. And a manufacturing manager who has run a real production line under real regulation, because the day we make something people put on their bodies, I want the adult in the room to have scar tissue.

DevOps researchers, too, if we can find them. And we will hire the right way: real contracts, real pay, employment law read before the offer, no one “brought over” on promises. I mention it because the alternatives exist in this industry, and the point of building a company with the word sapiens in it is to not do that.

Sales

I’m also the sales guy. If you want the DECK hardware, the software, or the research group’s time, the business email on our contact page is the door, and yes, it may start sending sales mail as well as receipts. If that’s not what you signed up for, say so and I’ll fix it — that’s the consent rule, and it applies to me too.

Architecture of Homo Sapiens sponsored this message. So did Claude and Codex, in the sense that they did most of the typing.