I'm playing with a new UX for working with agents in Biscuit, inspired by Claude Code, OpenClaw and Linear.
You have one central chat, which doesn't do work. It can only delegate work by creating tasks.
Tasks are plans, but with status, comments & activity. It has an orchestrator that does discovery on the task, and spawns a team of agents to build it.
Once its built, we automatically run QA on the result, and finally mark the task as "review" and hand it back to the user.
I think this makes multitasking, async and proactive tasks much easier for us. E.g. we spawn a task when certain threshold of bug reports come in to fix it.