The Agents panel
It opens on its own the moment the session’s first subagent starts — as long as the session is in the active workspace. Missed that, or closed it since? The Agents button in an agent session’s toolbar opens it by hand. The orchestrator — the session itself — sits at the root of the tree. Under it, each subagent shows:
Clicking a subagent focuses its transcript in the split — reopening the split too, if you had closed it.
The transcript split
1
Opens on the first subagent
The session pane splits: the terminal on one side, the transcript viewer on the other.
2
Streams the work live
The subagent’s own text, every tool it calls — one compact line, with the tool’s name and target — and each call’s result.
3
Scroll follows the tail
Like
tail -f: new output pushes the view down. Scroll up to read something that passed, and it stops following — until you’re back at the bottom, where auto-scroll picks up again on its own.4
The summary gets highlighted on completion
When the subagent finishes, what it did shows up highlighted right in the transcript.
Firing off a subagent doesn’t ask for approval. What it does afterward is what counts: every command the subagent runs goes through the same approvals inbox as the session, with the same risk classification — green passes on its own, yellow and red wait for you.
Subagents exist in Claude Code sessions. Codex sessions don’t spawn subagents — the Agents panel stays empty in them.
See also
Claude Code and Codex
Where an agent session is born.
Risk classification
What lands in each inbox color.