How a conversation id resolves to its running agent.

Two modes, chosen with config :agentix, :addressing:

• :local (default) — a node-local Registry. An agent is reachable only from the node running it.
• :global — Erlang's :global name registry, so a conversation resolves the same way from every node in the cluster.

Which one you need

:local is correct for a single node and wrong for more than one, in a way that does not announce itself. A conversation is meant to have exactly one agent, because that agent is the single writer of its durable log. Under :local on two nodes, each node's ensure_started/2 finds nothing locally and starts its own, so two agents interleave writes to the same conversation. Cancelling has the mirror problem: the cancel resolves nothing on the calling node and returns :ok having done nothing at all.

Under :global the name is cluster-wide, so the second node's start loses with {:error, {:already_started, pid}} — which Agentix.Conversation already handles — and every entry verb reaches the one live agent wherever it runs.

What :global costs

Registration is a cluster-wide synchronous operation, so starting a conversation is slower and gets slower as nodes are added. It suits workloads where conversations are coarse (a chat session, an extraction run) rather than one per request.

A netsplit is the case :global does not solve. Both sides register their own agent, both write, and on heal :global resolves the duplicate name by killing one — after both have written. If that matters, the durable log is the place to settle it, not the registry.