Field notes, essays, and further reading.
Everything we have published on specialist agents, the operating model, and the institutional discipline that makes them trustworthy.
We prefer to compete on substance. Nothing in this library is proprietary; the moat is in the operating system, not the essays. Read at your own pace, in any order.
Canonical chapters
The thesis
Five claims about specialist agents, read once.
The vision
The full argument. Market position, moat, and operating model.
The method
Seven pillars, in compounding order.
Modelsmith
The product. Evaluation-first specialisation engine.
The roadmap
Five phases from wedge to fleet.
Method
On eval design
Governed scenarios you approve. Expansion scenarios the loop proposes. The separation is what lets the autonomous iterate loop run without drifting its own rubrics.
On latency
In programmatic advertising the decision window is measured in tens of milliseconds. Specialist SLMs change what is feasible; a generalist model missing the budget is not slow, it is inoperative.
Operating model
On the fleet
A fleet of specialists implies a fleet of training hosts. The router is deterministic, not learned. The moat is the specialists, not the dispatch.
On promotion discipline
Six explicit states from candidate to production-accepted. Approval gates at the transitions that matter. Modelsmith supplies the artefact and the rollback contract; you control the shadow and the canary.
Field notes
On build vs buy
Most teams do not fail because they cannot run a fine-tune. They fail because they cannot institutionalise the closed eval–train loop and the promotion discipline around it.
The fork workflow
Platform improvements flow upstream. Customer-owned domain artefacts stay in the customer environment. The boundary is a deliberate design choice, not a legal formality.
The archive grows with the platform. Each persistent failure in the iterate loop, each decision with enough generality to be worth writing down, eventually ends up here.