The Two Seconds at the Door

What does it know? The written rules sit in policy libraries and compliance systems. The unwritten rules, the ambient state, and the judgment reside overwhelmingly in the collaboration layer: in email, in shared drives, in messaging tools. Those are the places no one lists when asked to name the firm's systems of record, and they hold the most valuable knowledge the firm possesses.

Similarity retrieval answers the question of what looks relevant. A graph answers the question of what is actually connected. The first fails when the decisive fact bears no textual resemblance to the signal.

Memory is what separates an agent that is merely competent from one that is continuous. An agent without it greets every returning user as a stranger.

What is it permitted to see, and on whose behalf? The person at the door looks through the glass before opening because not every caller is admitted and not every room is open to every visitor.

When a firm assembles its library by having a model read, summarize, and cross-reference its documents, a fact extracted from a restricted source and written into a summary loses the access control that governed the original. The barrier applied to the document. The derived summary inherited nothing.

A firm that builds its knowledge base without engineering for this has constructed an efficient mechanism for making restricted information broadly legible — and it has done so without anyone deciding to.

Has it brought the right context and the right participants to bear, and only those? Gathering the relevant context and deciding whom to involve are not sequential steps. They are one judgment.

Executives tend to assume that more context, and more agents, must produce a better result. The assumption is mistaken, and setting it aside is much of the skill. The person at the door does not summon the entire household for every caller. They convene the smallest sufficient response.

The graph is the firm's address book of relationships. The connector — of which the Model Context Protocol has become the common standard — is the telephone line it uses to place the call.

Is what it knows still true, and where did each piece come from? Context has a shelf life. An agent acting on a fact that was accurate last quarter can fail as badly as one acting on no fact at all. A capability records not only what is true but when it became true and whether it still holds.

Provenance has a second part, discussed far less often and mattering more. Agents do not only consume context. They produce it. A mistaken conclusion, recorded once, can harden into an apparent fact and propagate through the system.

Governing what an agent may see has a counterpart of equal importance and far less scrutiny: governing what it is permitted to write back, and what it is permitted to teach.