FWIW, I think the right architecture for both the in-memory and distributed coordinator is to consider an individual client connection as a special case of the more general "multiagent" idea. That is, a normal end user client is a multiagent where the number of agents is exactly one.

It's much easier to maintain and understand code that computes the general case, and then at the edge, we have adaptations the wire up clients with exactly one agent, rather than trying to build two different sets of logic and keep them consistent.

So, in the core logic of the coordinator, clients are all multi-agent, and we compute and queue up updates on them based on the set of agents they've asked to connect to.

Then, at the very edge we have special case code that either serializes the nodes over the websocket, or sends them out over channels/callbacks depending on whether we have an end user, remote client, or a coderd, in-process client.