cmux workspaces are underrated for this. The visibility problem of knowing which agent is doing what without constantly checking is real, and most people try to solve it with more logging when the actual bottleneck is the mental model for tracking state across sessions.

The Obsidian dashboard approach makes sense but I'd be curious how you handle agents that go silent. Not crashed, just stopped mid-task without raising an error. That was the hardest failure mode to catch on a multi-agent setup because there's no exception thrown, just missing output. Heartbeat files helped but only after burning a couple of sessions waiting for work that was never going to finish.