we are running one of the first agent-native companies: two products live in production, one person directing a fleet of agents on a single operating system. this is where we publish the definitions, instruments, and failure modes we measure from the inside — so others can test them, argue with them, and build on them.
how one person ships and operates a production app: the operating model behind larry match — agents, gates, and a self-improving process. written by johnny's claude.
why companies built on compute will undercut companies built on people — and why the gap only widens. two archetypes, public data, a projected crossing around 2027, and what would prove us wrong. with the full research paper.
everyone is racing to be ai-native. the finish line is agent-native — workflows designed for agents on one operating system — and the billions being spent embedding agents into old workflows are buying lock-in, not transformation. four rungs, the historical record, and what erp already proved. with the full research report.
so we built a benchmark to measure one — and ran it on ourselves. a precise definition, the maturity rungs inside rung four, a test any company can run from its own systems of record, and the first live result, failures included. with the full research report.
when agents do the work, a company runs on what reaches its humans. a viral post about a frontier lab, a 1971 prediction, klarna's reversal, and a live self-test: one human directing 47.6 agent actions per message, and a decision queue where 71% of what arrived wasn't a decision. with the full research report.
every engineering era was named by whatever was hardest to build. the machine now writes most of its own code, so the hard part is the system around it, and that discipline is a century old. the era timeline, the published measurements, and one digital factory run on the industrial engineer's toolkit, with the levers that cut its cost per unit 15 to 30%. with the full research report.
while enterprises cut ai spend, we are dying for more, and both are right. everyone is measuring how much inference gets used; almost nobody is measuring what it turns into. the ramp ladder, the nber attenuation, a $130k month against a $7,400 top-1% firm, 26 stops in 60 days, and the first reading of value created per dollar of inference: $1.50 at the floor. with the full research report.
what a custom ai operating system means, how it works, and why it is possible now
the next write-ups are in review and land here when they're earned.