Claude Code subagents: how to see what each one is doing

Where Claude Code shows running subagents

Ctrl+T is a different view: Claude's own checklist, not the subagents. If that one is blank, see Claude Code task list empty.

What disappears when a subagent finishes

• A subagent that finishes successfully leaves the panel at once. /tasks keeps it, marked done, for 30 seconds, and the footer shows /tasks to see subagents for those 30 seconds.
• A subagent that fails or that you stop keeps its panel row for 30 seconds and leaves /tasks.
• Results reach the main conversation as a completion notification in a later turn. Ask about progress before then and Claude says the subagent is still running.
• Each subagent's transcript is saved in a subagents/ folder next to the session's transcript under ~/.claude/projects/, so the detail survives even when the row is gone.

Log every subagent with a hook

SubagentStart and SubagentStop hooks fire when a subagent begins and ends. SubagentStop receives agent_type, agent_id, agent_transcript_path and last_assistant_message (the subagent's final text) on stdin, so one command keeps a log of every run:

{
  "hooks": {
    "SubagentStop": [
      {
        "hooks": [
          { "type": "command", "command": "jq -c '{at: (now | todate), agent: .agent_type, id: .agent_id, transcript: .agent_transcript_path}' >> ~/.claude/subagents.log" }
        ]
      }
    ]
  }
}

A subagent that reports through the SubagentHandback tool (v2.1.271 or later) delivers its report in that tool call, and last_assistant_message then holds only its closing text.

Limits worth knowing

• Up to 20 subagents run at once by default. Another spawn fails with Concurrent subagent limit reached; CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS changes the limit.
• Background subagents send their permission prompts to your main session and name the subagent asking. An answer that lasts beyond one call, such as for the rest of the session, applies to the whole session. Waiting for input without an alert covers getting told about those prompts.
• Each subagent starts with a fresh context: its own prompt, the task message Claude writes, your CLAUDE.md files and a git status snapshot. It doesn't see your conversation, and the built-in Explore and Plan agents skip CLAUDE.md.

A step per subagent on Taskhold's live board

Subagents inherit the main conversation's MCP tools, and background subagents keep every MCP tool, so a subagent can call Taskhold's tools the same way the main session does. The pattern: the main session plans the task into steps with taskhold_plan, one step per subagent, and puts the step's id in each delegation message. Each subagent reports taskhold_progress on its step and completes it with a summary. The live board then shows one bar split by step, which steps are running, what each subagent last said, and any question one of them blocked on.

## Subagents and Taskhold
- Before delegating, taskhold_plan the task with one step per subagent and
  put that step's id in the subagent's prompt.
- Subagent: taskhold_progress on your step at milestones (N of M plus one
  line), then taskhold_complete it with a one-line summary.
- A decision only the person can make: taskhold_block the step with one
  question.

• A subagent definition with a tools list gets only the MCP tools it names. Add mcp__taskhold__* to the list, or leave tools out.
• Explore and Plan skip CLAUDE.md, so put the step id and the reporting rule in the delegation message itself.
• Progress is what each subagent reports. The card's status dot shows when an agent last reported, so a silent one stands out. If the tools don't load at all, see MCP server not connecting and Taskhold troubleshooting.