Actors

A human or an agent, and the practice never presumes which. An actor may hold a role if — and only if — it has what that role requires: eligibility turns on access, not on actor kind. There is no presumption that a human is the better actor, or the default one.

Every actor must be distinctly identifiable — a human by email or username, an agent by model version, skill and unique id. Identity is not bureaucracy; it is what makes accountability traceable.

The consequence is that the practice runs all-human, all-agent, or any mix, without alteration. It is also why reachability of information is not a later concern — it decides who is eligible to act at all, and what a role requires an actor to reach is written there.

What the actor model is for

1. The whole practice can be run by humans only.
2. The whole practice can be run by agents only.
3. It can run on any mix.
4. Every agent actor holds its role by permission delegated from a human.

The fourth is what stops the second being a human-free practice.

The four properties are not symmetric, and nobody expects that

The model is mechanically enforceable against agent actors and self-declared against human ones. You can withhold the working session from an agent; you cannot un-show it to a human. You can supply an agent's references and check what it was given; you cannot verify that a human read the standard.

So the all-agent property is verifiable and the all-human property is true but unverifiable — the reverse of what most readers assume.