Background subagents give the user zero visibility; model waits silently for 15+ minutes

Summary

When Claude Code launches background subagents (Agent tool) for a long research/drafting task, the user gets no progress at all until the agents finish. In my session the main agent set up a task, spawned two background subagents, printed one line, and then produced nothing for ~18 minutes. I had no way to see what they were doing, whether they had stalled, or whether the premises the main agent had chosen on its own were wrong.

What I expected

• The main agent should state its premises (what it decided, what it assumed) before delegating, so I can correct early.
• While background agents run, some periodic status: which agent is on what step, what is slow, rough ETA.
• Ideally the UI itself surfaces background-agent activity (last tool call, elapsed time) without me having to ask.

What happened

• Main agent (Fable 5.1) decided the scope alone, spawned two background Agents, and ended its turn with a one-line "waiting for drafts".
• ~15 minutes of silence. I asked for status; only then did it check the agents' transcripts and report that one of them was blocked on a series of slow serial CLI calls.
• Only after I complained again did it lay out the decisions it had made so I could check them.

Repro

1. Ask for a multi-part deliverable that needs several minutes of subagent research (e.g. many slow CLI/DB queries).
2. Model spawns background Agents and ends its turn.
3. No output until completion notifications arrive.

Environment

• Claude Code CLI, macOS (Darwin 27), model claude-fable-5-1
• 2026-09-21

Ask

1. Model behavior: surface premises before delegating; post interim progress while background agents run; proactively offer to cut scope when one agent is the bottleneck.
2. Product: show background-agent activity (current tool call, elapsed time) in the main session UI, and let the user interrupt/redirect a specific agent without waiting for it to finish.

Start by reproducing the multi-part CLI workflow described in the Repro section with background Agents and slow CLI or database queries. Trace how the main session reports delegated work and how completion notifications are surfaced. Done means users can see premises before delegation, receive interim activity or progress, and inspect or control a specific background agent without waiting.