One studio. One brain. Agents with jobs.
We run this company on an operating system we built: one shared brain, agents with jobs, humans making the calls. These notes document the experiment in public.
Does an AI-native company still need departments? What replaces the meeting? Essays from inside the experiment — written as we learn, not after we've won.
The thesis
The AI-native company
Series 01, essay 01: what changes when a studio runs on an AI operating system — one shared brain, agents with jobs, humans making the calls.
Read ›Org design
Does an AI-native company still need departments?
Series 01, essay 02: a department is a standing promise to keep paying attention to a domain. We deleted the org chart and wrote a schedule instead — and here is what that traded away.
Read ›Prior art
We went looking for someone already doing this.
Series 01, essay 03: we scanned the field for a company already run by one agentic brain, hoping to copy it. What we found was a lot of tools, a few demos, and one empty quadrant.
Read ›Verification
Don't let the agent grade itself.
Series 01, essay 04: 2026's agent bottleneck isn't generation, it's verification — and the deepest failure is self-evaluation. Unless the evaluator is isolated from the evaluated, no score can be trusted. Here is the structure we run instead.
Read ›Harness
The harness has an expiration date. Ours mostly doesn't.
Series 01, essay 05: as models get stronger, the bitter lesson says delete your scaffolding. True — for the harness that compensates for a weak model. But a second harness (memory, accountability, permission, cadence) compounds instead of expiring. Here's how we tell them apart.
Read ›Design law
Build the floor, not the ceiling.
Series 01, essay 06: in 2026 guardrails became the bottleneck. But a gate is either a floor (guarantees an outcome, scales with the model) or a ceiling (restricts a process, expires). Build floors; keep only the irreversible as a thin membrane. The design law that closes the harness trilogy.
Read ›The shift
The bottleneck moved.
Series 01, essay 09: a frontier lab just co-founded a $1.5B services firm to deploy its own models. That is a price signal: the scarce thing in AI is no longer the model. It is everything a company knows that its software does not.
Read ›The category
YC put a name on the company brain.
Series 01, essay 10: Y Combinator made the company brain an official startup category and has funded at least five of them. We have been running one since June. Notes on what the category will get wrong, from inside the thing it describes.
Read ›The graveyard
Apps for meeting nearby keep dying.
Series 01, essay 11: the proximity graveyard is the most useful document in the category. Read separately, the autopsies show three doors: trust, density, consent. None of them is locked. Why our flagship is built to walk past all three, and what would still have to go wrong.
Read ›In progress — 07 · What replaces the meeting?