Epistemic just means “having to do with knowledge.” It’s also a word that scares people off, and I understand why, because it scared me first. I spent my formative years trying to read philosophy, and words like ontology and epistemic have carried a built-in fear for me ever since.

When I was a teenager, my older brother went off to high school and left his books behind. German philosophy — Hegel, Kant, Heidegger. I tried to read them for years and understood roughly one paragraph in ten.

Then I went to college and lived with philosophy majors. At two in the morning somebody would ask whether we could really know anything, and the argument would go until the sun came up: Chomsky on how we know so much from so little evidence, Rorty on whether the mind mirrors the world. The vocabulary stuck. The confidence never arrived. So whenever somebody starts talking about ontologies in a meeting, history comes along with it.

The definition I should have had at fifteen: epistemic means “of or relating to knowledge or knowing.” That’s Merriam-Webster, in full. An epistemic question is a question about what you know and how well you know it. The noun behind it is epistēmē — Greek for knowledge, from a verb that means “to stand.” To know something is to have somewhere to stand.

One more fact, because it surprised me: the word is younger than it sounds. Its first known use in English is 1922. Kant and Hegel never used it — it was built for the modern study of knowledge, long after they were gone.

The word has moved since. Until recently, epistemic mostly lived in philosophy and machine-learning papers, where epistemic uncertainty names the not-knowing that more data can fix. In the last two years it has jumped to the center of a practical question in software: when an AI agent acts on your behalf, what does it believe, and is any of it true?

I run into that question constantly, because it’s what my last two papers are about — The Layer Nobody Governs and The Other Permission System. Between them is the idea I call semantic and epistemic governance. The semantic layer is what things mean to a system. The epistemic layer is what that system takes to be true: what it can know, how well, and what it’s allowed to operate on.

Every agent platform on the market governs what an agent can do — which tools it may call, which files it may read, which actions it may take. That’s the action layer. NVIDIA shipped a hardware watchdog that quarantines an agent in milliseconds. Progent cut the success rate of tool-abuse attacks from 41 percent to 2. OWASP published its first Top 10 for agentic apps. The action layer has real security now.

Almost nothing governs what an agent knows. That’s the epistemic layer — and when I use the word there I mean something specific: the set of facts an agent holds as true, the premises it operates on, the conclusions it’s licensed to draw from them.

Epistemic governance means building checks and limits for that layer — governing what a system is allowed to treat as knowledge. Is this fact in memory planted or earned? Does this conclusion actually follow from its premises? Can the agent tell “I remembered this” apart from “I figured this out”? A locked-down tool set doesn’t catch any of this — a clean action door is just as happy to act on beliefs that shouldn’t be there.

If that sounds abstract, the attack numbers say otherwise. Agents write to their own memory, so attackers can get the agent itself to record false facts and come back for them later — success rates run past 80 percent in the research. Agents talk to agents, and a false premise propagates from one to the next, shaping every action after.

When an agent breaks a rule on the action layer, alarms go off — quarantine, block, alert. When it’s wrong on the epistemic layer, nothing goes off: it produces a coherent, confident report, correctly reasoned from a premise nobody checked, and no log line says what went wrong. You find out when the damage shows up somewhere else.

So: epistemic means having to do with knowledge, and epistemic governance means pointing the same effort we’ve spent on governing agent actions at agent beliefs. You can restrict everything an agent does and still get burned by what it believes.

That’s the whole word. It’s an adjective about knowledge — where knowledge comes from, how well you have it, what you’re allowed to treat as true. The definition turned out to be the easy part. Deciding what our agents are allowed to believe is the work.