Post

wait does the multi-agent version actually route better or does it just feel cleaner architecturally? every time I split agents by tool specialty the coordination overhead eats whatever I saved

That's a great question! But please remember that both implementations are just toys, so it's hard to judge. Furthermore, I don't really have a harness of evaluation to prove one is better than the other. However, allow me a couple remarks.

I had one situation where my agent was going back and forth between different tools of different agents. It was doing a mix of bash and websearch. I was using the multiagent variant. And it seemed the handoff was working fine, although at first I thought it'd struggle to go back and forth.

In multi-agent setups I've had better results when each agent gets a tiny tool surface plus a shared durable task state. The failure mode is usually handoff loops, not raw model quality.