Skip to main content

The Lab · The Agentic Operating Model

The agentic operating model: propose broadly, dispose certainly

One founder and a gated fleet, run like the product is built: a council decides, builders build, a Haiku fleet sweeps in parallel, and the gate verifies everything before it is trusted. The receipts: 13 versions in 48 hours, a 50k-line AI-written OS, 327K lines in four days — with a conflict-free chapter factory by construction.

JR
Jon RossFounder, vocabotics — 15 years building safety-critical systemsResearch write-up · 2 min read · reviewed 2 July 2026
Verified by a human. Drafted with AI, verified by a human. Jon Ross, 2 Jul 2026
Living document. Reviewed 2 Jul 2026
Shipped & measuredDownload PDFPDF · 773 KB

The company is run the way the product is built: propose broadly, dispose certainly.

13 / 48h
versions shipped
50k lines
AI-written OS kernel, 0 errors
327K
lines in four days
0
unverified artefacts trusted

The most-copied idea of the last two years is "agents". The part almost everyone skips is the discipline that keeps a swarm of agents from producing a swarm of confident mistakes. Our answer is an operating model where the company is run the way the product is built: propose broadly, dispose certainly.

The four layers

Shipped & measured
  • The Council (Opus) decides and ratifies — unanimously, or it fails.
  • The Builders (Sonnet / Opus) write the hard code: memory-safe, floors green or revert.
  • The Fleet (Haiku) sweeps the bulk work in parallel — gather, batch, generate — in minutes rather than days.
  • The Gate trusts nothing an agent makes; it verifies it.
  • The Dashboard is shared, verified memory — no private silo survives.

The two-engine thesis is the whole of it: propose broadly with the fleet, then dispose certainly with the gate. Generation is cheap and wide; acceptance is strict and narrow.

The receipts

Shipped & measured

Tempo you can point at: 🟢

  • 13 versions in 48 hours.
  • A 50,000-line operating-system kernel, AI-written, with zero errors (dated November 2024 — well ahead of the "AI writes real systems" discourse).
  • 327,000 lines in four days during the PRISM sprint.
  • A chapter factory that is conflict-free by construction — parallel agents that cannot collide, because the structure forbids it rather than the process discouraging it.

Why one founder can out-discipline a mega-lab

Not by working harder — by subtracting. The character proof is the cull: 57 projects archived in a single day to close a membrane and ship one governed world. You can copy an org chart; you cannot fake having had the discipline to delete most of your own work to finish the rest. The operating model turns one person plus a gated fleet into something that proposes like a large team and accepts like a safety assessor.

Why it ladders back

If one person can do what took a thousand — and every output is verified before it counts — then capability compounds without the usual tax of coordination and error. That is abundance, demonstrated on ourselves before we sell it to anyone.

The open edges

Kept visible on purpose. These are what turn 🔬 into 🟢 — honestly.

  • Tempo figures (13 versions/48h, 327K lines/4 days) are throughput receipts, not a quality claim on their own — the gate is what makes them safe.
  • The model depends on the gate holding; a fleet without a gate is just faster mistakes.

Sources

  1. vocabotics Dashboard — Cross-cutting: The Agentic Operating Model, 2026-07-02vocabotics internal record · as of July 2026

    We use cookies.