Agents

A process and a record

An agent is two things at once: a process on the user's machine and a resource — the record linking the real machine to that process. Declaring an agent never creates a machine: the record waits for its agent to connect. The machine opens no ports; the agent connects outward, with its own token.

How the agent gets onto the machine — two equal paths that converge into the same connection:

Both paths run one and the same install script — two scripts would drift.

Capabilities are written, never discovered

What a machine CAN do is written onto its record by a publisher — an installer, a person, a machine image. A capability has a name, labels, an informative version, and a readiness flag. It belongs to the machine, not to whoever published it:

Readiness is one formula

An agent is ready when two conditions hold at once: the agent is connected, and every need of the declaration is met — the required capability is on the record, ready, and matches the label constraints. Ready on the agent's handle waits for both: the refusal comes before work is dispatched to the machine, not after it fails there.

Needs match capabilities by name plus label constraints — equality and "one of". Versions are never compared.

Selection and fan-out

Agents can be selected without knowing their names: a selector over labels and needs returns a snapshot of the matching agents; an action addressed to the set executes on every machine of the set.

Foreign agents

An agent declared by someone else can be attached: recognized and read like any handle, never created. An attached agent takes actions like your own; the only difference — it cannot be a parent or a child in the ownership tree.