Context as the moat

Code, configs, architecture decision records and docs are all prompt material. That makes architectural legibility and stack uniformity attention-efficiency measures rather than aesthetic preferences: six observability stacks in one repository means the model burns attention working out which one is canonical.

A 2022 chat thread, an undocumented infrastructure module, a convention everyone senior simply knows — none of it exists as far as an agent is concerned. The response is a checked-in system of record: skills, context, recipes, decision records, so agents understand the system before they change it. The goal is not better documentation; it is agents that can operate safely.

When the binding constraint lives outside the repository — a design system, a cost decision, a product ruling made in chat — every in-harness gate reports green and the result is still wrong. An agent handed a design file used a generic component instead of the design system’s export, which encodes accessibility and async behaviour; ticket, PR and skill evals were all green, and the failure surfaced downstream as a failed accessibility audit.

Also: Figma Code Connect and the Figma MCP server; Linear; Notion; Granola.

The brief format, the registers, the role skills, the findings corpus and the accumulated guardrail history are precisely the “not the model” investment — built on the wager that harnesses and models turn over and the context does not. Netlify’s framing is a sharper mission statement than Assay’s own docs: not better documentation, but agents that can operate safely.

Assay ships the methodology as an open, installable layer and keeps the operator’s own material — the trust roster, the findings, the accumulated rulings, the rationale behind each guardrail — on the operator’s side. Read against this evidence, that is this concept restated from the buyer’s side: the published bundle is deliberately the commodity layer, and the part that cannot be bought is the part that stays local. A defensible position, not an accident of what happened to be ready to publish.

A ruling or a priority that lives in a session channel rather than in an issue, a brief or a findings entry is exactly the out-of-repo constraint that turns every green gate into a false negative. Two mechanisms are designed and unbuilt: a field on a brief naming the out-of-repo artifacts it depends on, and a dispatch gate that refuses to claim a brief whose linked rulings are still provisional or open. The expert-usage crawl — mining which files an expert opens, in what order, for which class of work — is a mechanism the dreaming pass does not have and could copy. And “chat is for exceptions, and resolved exceptions become new signals” is the findings lifecycle described from outside.

• Whether anyone measures the moat rather than asserting it: a same-suite model swap with quality held constant would be the first hard number behind “evals are the durable asset”.
• Whether harness vendors ship a way to version out-of-repo context (design systems, product docs) so evals stop rotting silently against third-party state.
• Whether the expert-usage-crawl pattern appears anywhere outside Microsoft — it is the only mechanism in this scan for capturing know-how nobody wrote down.