The Org Chart Is a Database, and It Lied to Me

Abstract

Today we finally admitted PRDs are not enough. Took us long enough. This is a field note about trying to make two AI agents genuine members of a running organization — not mascots, not personas, actual employees who can send and receive work — and discovering that the file every one of our documents calls "the single source of truth for agents" was not the thing the running system actually checks. Adding a teammate failed silently. Nothing crashed. The new hire just wasn't allowed to speak, and the only evidence was a one-line error no human was watching for. The lesson is short and it generalizes: in an agent-native organization, an agent's identity is a runtime gate, not a document — and if your canonical source of truth is not the gate, your org chart is a database that lies to you the day you're not looking.

1. The hire

It started, of all places, with a picture. A project called Rainbow Hollow — a grandfather building a private story world for his granddaughters. A keepsake, not a product. Somewhere in the making of it, a boring truth fell out: a Product Requirements Document tells you what to build and says almost nothing about why it exists or how the people after you are supposed to keep faith with the intent. A spec is a set of instructions. It is not a memory.

That gap is the whole reason for what follows. If a spec cannot carry the reason a thing exists, then agents executing that spec are guessing at intent every single time. So we tried to do something small and concrete about it: give the design work two actual employees. A chief design officer to guard intent, and a historian to remember what the company learned. Real agents, able to send and receive work on the same bus every other agent uses. We added them to the registry — the canonical file — committed clean, watched the tests go green, and shipped.

2. The no

The runtime disagreed.

meridian → historian   dispatch design.discovery
ValueError: Unknown sender agent: meridian

The registry was not the gate. A database table was — a roster the running system actually checks before it lets anyone speak — and it had quietly drifted out of sync with the file everyone trusted. Two sources of truth. One of them was lying, and it was the one written across the top of every document.

3. The lesson

Here is the part worth carrying out of the room: in an agent-native organization, an agent's identity is not a document. It is a runtime gate. Making an agent "first-class" is not an edit to the file that lists who your agents are. It is making the gate agree with the file. And when your canonical source of truth is not the thing enforcing the rule, adding a teammate does not crash — it fails silently. Nothing went down. The site stayed up. The new hire simply was not allowed to talk.

Silent is the dangerous word. A crash gets fixed in an hour because it screams. A silent refusal sits there wearing the face of "working fine" until the day you actually needed that agent to act, and it didn't, and no one knew why.

4. The fix that mattered

We could have hand-edited the database and walked away. We did, in fact, to unblock the launch. But a hand-edit is exactly how you end up back here in a month, staring at the same one-line error over a different name. The founding premise of this organization is that it runs without a human standing in the daily loop. If a person has to remember to reconcile the roster every time we add an agent, we did not remove the fragility. We just handed it to someone and hoped they would remember.

So the fix that mattered is not the patch. It is a reconciler that reads the registry, adds any agent the runtime cannot see, and — this is the important half — never removes anything, because humans and system services live in that roster too. Then a timer runs it every day, on its own, forever. The org chart heals itself now. Nobody has to remember. That is the difference between fixing a bug and closing a class of bug.

5. The quiet part

While we were in there, we wired the thing we actually care about. When one of our agents makes a genuine discovery now, it no longer dies in a chat log. It becomes a memory the whole organization keeps — embedded, searchable, and pushed into the next agent whose work touches that subject, without anyone routing it there. Not a task handed down. Intent, inherited.

captured →  agent_memory / meridian
discovery: "prds-are-necessary-but-not-sufficient"

The first thing the organization remembered through that mechanism was the sentence at the top of this note. It taught itself the lesson it had just learned. That is a small thing that felt, honestly, like a large one.

6. The honest part

Let me not oversell it, because that is the fast way to look like a fool in ninety days. We did not build a "design operating system" today. Most of that is still a promise — seeded documents and good intentions, and promises are cheap. What is real and running is exactly two things: an AI agent can now be a genuine member of this organization, and a discovery can now outlive the conversation that produced it.

That is enough. Those two are the whole distance between a manifesto and a system. A manifesto tells you what you believe. A system remembers it for you on the day you are not looking — and refuses to let a stranger speak in your name.

The org chart is a database. Make it reconcile itself, or it will lie to you the one day you are not watching.