Keep investing in live, persistent, communicative sub-agents (tmux-style panes) — it changed how I work

Bug Description
This is a feature-request / encouragement issue, not a bug. I just ran a long session where a "team-lead" Claude orchestrated several persistent sub-agents, and the experience was so much better than opaque one-shot tasks that I want to flag exactly what worked so it gets preserved and expanded.

What I loved, concretely:

• I could watch each agent work in its own terminal pane (tmux-style splits). Over one session the lead kept multiple named agents alive — one doing the visual/animation work, one doing mobile/PWA/SEO, plus short-lived specialists for isolated fixes — and I could see each one's reasoning and commands live, not just a final diff.
• Persistence across the whole arc mattered. The same agents picked up loop after loop (cards → sound → mobile → PWA → SEO → perf), ~50 merged PRs, carrying context forward instead of re-explaining from scratch each time.
• Synchronous relay of findings is the killer feature. The lead surfaced each agent's results to me as they happened, so I could steer mid-stream — e.g. when an agent reproduced a stubborn mobile bug (a video showing a play button instead of autoplaying), I watched it isolate the real root cause (a cold-fetch race, not the obvious muted-attribute theory) by driving Chrome under a strict autoplay policy. I'd never have trusted the fix without seeing that work.
• Visibility made failures diagnosable. When the browser-automation MCP died mid-session, I could see an agent adapt by driving Chrome directly over CDP and keep verifying. Watching it recover was reassuring in a way a silent retry never is.

Two things that would make this even better (the rough edges I hit):

1. Faster/automatic agent pickup. Background agents reliably went idle right after being assigned a task and needed a manual "nudge" to actually start — the lead spent real effort babysitting this. Reliable wake-on-assignment would remove most of the coordination overhead.
2. First-class per-agent worktree isolation. Two agents operating in the same git working tree collided once — a duplicate PR and a HEAD switched out from under one of them. Giving each agent its own worktree by default would let them run in parallel safely, which is the natural next step for the live-pane model.

Net: the visible + persistent + communicative pattern is exactly what I want from agentic coding. Please keep building
toward it.

Environment Info

• Platform: darwin
• Terminal: iTerm.app
• Version: 2.1.185
• Feedback ID: 4f56c278-37ba-4a06-b3f2-85f6f742a0d2

Errors

[]