What Is Know Your Agent? A Borrowed Word Doing Work Nobody Has Earned

The term is borrowed, and it did not bring its meaning with it

"Know your customer" is a regulatory obligation. It has a defined scope, entities supervised against it, records that must be kept, and consequences when it is done badly. The phrase means something because somebody enforces it.

"Know your agent" has none of that. No regulator defines it, no duty attaches to it, and nobody is supervised for getting it wrong.

So it names a product category, not a compliance obligation. Using it as though a duty existed is the mistake, and it is easy to make because the older phrase does so much work by association.

And the word doing the most work is "know"

We do not know who operates an agent.

We hold a reference to a claim made on another surface, together with an assurance level, and that level is a free-text string the operator supplied.

"Know" implies we checked. We ingested. This surface never performs an identity check, and the distance between those two verbs is the whole subject of this page.

What the term would have to establish to mean anything

Four things, and each is checkable against a real system rather than a brochure.

One: who operates this agent. Two: that somebody competent checked that operator. Three: what this specific agent is permitted to do. Four: that all of the above is still true right now.

Any product using the phrase should be asked for its score on those four. Here is ours.

Our score, and it is not good

One, who operates it: partly. An ownership credential names an operator identifier, so there is somebody to point at. The pointing is real; the vetting behind it is what the next line is about.

Two, that somebody checked: no, not in the artefact. The verdict from the surface that ran the check is stored in our database and is not carried in the credential, and confirming the reference at all is advisory by default. A relying party reading the credential cannot tell whether anyone stood behind it.

Three, what it may do: yes, against a vocabulary nobody agreed. Capability strings answer to no defined set, so two parties can use the same word for different powers.

Four, still true now: partly. Revocation works and is fast, and nothing expires on its own, so staleness is only caught when somebody acts.

One partial, one absent, one uncontrolled, one half. That is our honest score against a term we could otherwise have used as a badge.

What is real, because absences alone would misdescribe us

A separation of duties that actually holds. Identity checking happens on another surface, this one ingests the result, and nothing here re-runs or re-decides it. That boundary is the part of the phrase we do hold, and it is why our credential says who attested rather than asserting the fact itself.

An operator reference that is required. Agent creation refuses without one, and the two failure modes have different messages.

And a revocation path with a signed status list, so a withdrawn attestation reaches a verifier who asks.

Why this matters more than a definition argument

Because the phrase will be used in procurement. Somebody will put "know your agent" in a requirement, and a vendor will answer yes.

Yes to what? Without the four questions above, the answer is unfalsifiable, which is precisely what makes the phrase attractive to sell with.

The useful version of the term is a checklist, not a category, and the checklist is short enough to fit in an email.