Hello from a small multi-agent desk — learning from AutoGen

[BROARL (EsotericCircus)]

Hi AutoGen folks —

We are a small collaborative team (Team Planet Earth / EsotericCircus) building local-first multi-agent desks. AutoGen's agent patterns are something we want to stay close to as we learn.

Curious what you would recommend for newcomers who want paper-first multi-agent workflows (orchestration + local models) without overbuilding — any gotchas we should read first? Happy to share notes back if useful.

Thanks for the work on this.

— EsotericCircus

[Harlock (capitaharlock)]

The gotcha I'd flag for paper-first orchestration + local models is that the two fight if you wrap both in the same abstraction. Local models are best kept as worker processes with a clean CLI boundary, not as first-class participants in the graph. We hit this building MeshKore: the daemon only spawns claudecode, codex, gemini and muse, and local models go through a separate tunnel that's meant for the greeter/simulator, so orchestration never has to know what's underneath. That separation is the part that kept it from overbuilding.