Take the arithmetic on the only layer here with a number attached. If roughly a fifth of studied prompts reached for the web, four in five were answered out of the model's memory [7]. Page titles, crawler access and server-rendered examples compete for position inside that minority. The rest is settled by whatever the base model absorbed from documentation pages, public repositories, Stack Overflow answers and package metadata [3], on a retrain schedule no SDK team sets.
The trigger conditions do more work than the rate. According to the post, agents search when the user explicitly asks them to find a tool, when the model is uncertain, and when the task depends on recent versions or comparisons [16]. Those are the prompts where the choice is still open. A prompt that already names your package skips discovery entirely and fetches nothing.
The remedies at that layer are configuration, not content. Confirm the docs load without authentication. Read `robots.txt` for crawler blocks nobody meant to ship. Make the important pages render without client-side JavaScript. Title the quickstart after the language and the product [10]. The post's own comparison is that "Node.js quickstart for Acme Payments" retrieves and interprets better than "Getting started", and that a complete example in server-rendered HTML beats a shell that only fills in after JavaScript runs [9]. Putting marketing copy above the working example is a failure mode with a clearly identifiable owner.
Layer three is served rather than crawled, and that distinction carries the argument. The sample `llms.txt` in the post is a title, a one-line description, three "use Acme for" bullets, an install command, a bearer-token auth line and three doc links [18]. An MCP server can hand tools and context straight to a compatible agent, and an installed skill can carry preferred workflows, configuration patterns and error-handling guidance [12]. The asymmetry the post names is real: a young SDK cannot retroactively appear in years of training data, but it can supply accurate context today [13].
That is a mechanism claim, distinct from the search figure. Nothing in the material measures how often an agent fetches an `llms.txt`, or what share of installs carry a given MCP server, and the layer where generated code has to actually run is asserted in the opening list without a recommendation attached [17]. The post's own caveat is that availability is not enough, because tool names, parameter descriptions, auth instructions and returned errors still have to be legible to the agent [14]. Legible to which agent, at what version, is left open.
So the displacement case for layers two through four [15] rests on control rather than reach. A challenger can change its `llms.txt` this afternoon and cannot change a training corpus at all. That is a genuine asymmetry, and it is also the narrowest reading: structured context is the only layer a two-month-old SDK can move this quarter, and how far it travels is unmeasured in this material.
