Why the company hasn't changed
David expected 2026 to look alien from where he stood in 2022. The models arrived. The org chart didn't move. He spent an evening trying to work out why.
David has a habit of standing in one year and interrogating himself from another. The version of him in 2022 was promised a transformed world: AIs that out-reason him on most things, close enough to AGI that you'd squint. That David would have told you the enterprise of 2026 would be unrecognizable. The David who actually arrived in 2026 looks around and finds the enterprise stubbornly the same. People spend more of the day inside a chat window. That's about it.
It's still being used like this tool that I talk to and I get a response from.
What gnaws at him isn't the gap between expectation and reality — it's that the obvious explanation feels wrong. Everyone reaches for model quality. The models just aren't there yet. David suspects that's a dodge. Freeze capability exactly where it is today, give the world five years to digest it, and the firm would still look radically different. So the lag isn't the model being undercooked. It's that we've been modeling "employee" as a fundamentally cognitive role, and it never was one.
His sharper move was to take the word apart. A firm, stripped down, isn't a machine for producing cognition. It's a machine for assigning accountability — a structure whose real job is to answer when this goes wrong, whose name is on it. A human employee is trusted not because they can do the work but because they can be held to it. They carry liability, reputation, a career that can end. An agent carries none of that. So the instant you let it own a function instead of assisting one, you tear open an accountability vacuum the firm has no way to absorb. Chat, in that light, is exactly the ceiling: it's the most you can extract from something you can't yet hold responsible.
The company is actually this orchestration of agents. Humans kind of set the direction, but the agents are actually the company.
That's the world he keeps picturing — humans as one input, the org as a dense weave of agents, everyone a "person" whether biological or synthetic. But the part he let himself sit with is that this might be the wrong attractor entirely. The Coasean reading runs the other way: firms exist because coordinating through the open market is expensive, so you internalize whatever's cheaper to keep in-house. Cheap, capable agents collapse coordination costs everywhere — including across the firm's own walls. The same force that could pack a company with agents also makes the company-as-container less necessary. The endpoint might not be the dense autonomous firm. It might be the firm thinning out, dissolving into fluid, agent-mediated coordination — capability spun up on demand rather than housed.
And the reason none of it has happened yet, he thinks, is that the missing piece isn't intelligence. It's a primitive that doesn't exist as infrastructure: persistent identity, memory, and attestable accountability bundled into something an organization can grant scoped trust to. Someone has to make an agent into a thing you can vouch for. Which is, of course, the exact seam David spends his days working in — a fact he notices with the faint embarrassment of a man who realizes the argument he just built happens to point straight at his own company.