No Ghosts in the Machine: Every Agent Should Have a Name

For most of computing history, "who did this?" had a short list of answers: a person, or a program a person ran. The accountability chain ended at a human, quickly. You could always walk it back to someone's hands on a keyboard.

That assumption is quietly breaking. When you run a mesh of AI agents — and we do — work gets done by things that aren't people and aren't quite "programs" in the old sense either. They make choices. They call tools. They act. And the moment you have a second one, you inherit a question the single-agent world let you ignore: which one of you just did that?

The ghost problem

Picture a system with a dozen agents helping run it. Something changes — a file gets written, a message goes out, a setting flips. If any of those agents could have done it, and none of them carries a verifiable identity, you don't have a system. You have a haunted house. Something moved the furniture and you'll never know which something.

We decided early that we would not live in a haunted house. The rule is simple to say and a discipline to keep: nothing acts in our systems without an identity we can verify. Every agent that can do anything real is registered, named, and accountable. If it can act, we can attribute. No anonymous helpers, no untracked processes quietly making decisions. No ghosts in the machine.

Why a name is a security control

This sounds like bookkeeping. It's actually one of the strongest security postures available to you, because it changes what an attacker has to defeat.

If your agents are anonymous, a hostile one — or a legitimate one that got hijacked — blends into the crowd. Its actions look like everyone's actions. But if every agent must present a verifiable identity to act, then the unidentified actor is immediately anomalous. It's not "one of many possible culprits." It's the one thing in the room without a name, and that alone is the alarm.

Identity turns "I think something's wrong" into "this specific actor did this specific thing." It turns forensics from archaeology into reading a log. And it makes the unaccountable agent — the exact thing you should fear most as AI gets more capable — the one thing your system is built to reject by default.

Attestation, not vibes

The other half is that an identity has to be earned and checked, not asserted. It's not enough for an agent to claim it's "the trusted one." A claim is just a sentence; anything can say a sentence. What matters is whether the claim can be verified against something the agent can't simply fake.

So we lean on attestation — an agent doesn't get to tell us who it is, it has to prove it, in a way we can independently confirm. The label follows the proof, not the confidence. (This is the same instinct that runs through everything we write here: a confident sentence is not evidence. Identity is no exception.)

Where this is going

I think this becomes one of the defining security problems of the next few years, and most people haven't hit it yet because they're still running one assistant, not a society of them. When you're running the society, "which one did that?" stops being a curiosity and becomes the whole game.

Our bet is that the systems that age well will be the ones where every actor has a name, every name is verifiable, and the nameless thing is treated as the threat it is. We don't claim to have finished this — it's a standard we hold ourselves to and keep tightening. But the principle we won't bend on is the one in the title. If it can act, it has a name. No ghosts.