Claude can hand parts of a task to subagents: one to search the code, one to write a test, while it carries on. Tori gives each subagent a lane, so you can read what it is doing on its own instead of seeing everything mixed into one transcript.
Lanes need the agent to report its subagents, which Claude does. Other agents say nothing about helpers, so their work shows in the transcript as the main agent’s own.
The lane strip
Once a chat has started a subagent, a strip of chips appears above the composer:
- main, the chat itself, always first. Its chip takes the agent’s colour while it works.
- One chip per subagent, named after what it was asked to do. Running ones come first.
Each subagent chip has a status dot and one figure:
| Figure | Means |
|---|---|
| 42s, 1m12s | It is running, for this long |
| 18k | It finished, using this many tokens |
| waiting | It is stopped on a prompt for you |
| Any other word | It ended another way, as the agent reports it, for example failed |
A finished subagent whose ending was never recorded, usually one that ran in the background in a session you reopened, shows as finished rather than failed.
Past two finished subagents, the rest fold behind a +3 done chip. Click it to show them all, and Hide done to fold them again. The lane you are reading always stays visible.
Read a lane
Click a chip, or press Option+1 for main and Option+2 through Option+9
for the chips after it, in the order they are drawn. The transcript narrows to
that subagent’s own rows: its tool calls with their names and arguments, its
output, its result.
Reading a lane changes only what you see. What you type always goes to the main agent, because there is no way to talk to a subagent directly. While you read a lane, the composer says so: Watching write tests. What you type goes to the main agent.
In the main transcript, the tool call that started a subagent has a Read what this subagent did button (is doing while it runs) that opens its lane.
When a subagent needs you
A subagent can ask for permission or ask a question while you are reading something else. So it does not stall unseen, its prompt is shown in the main transcript too, with Read what this subagent is doing to jump into its lane, and its chip reads waiting. Answer it from either place.
A question from a subagent is marked from a subagent, so it does not read as the main agent’s.
Background tasks
Work the agent started in the background, such as a long shell command, shows in a second group labelled Background, with how long each has been running. These are labels, not lanes: a shell task has no transcript of its own to read.
When the turn ends before the work does
A chat stays working while its subagents and background tasks are still going, even after the main reply has finished. The status strip says what it is waiting on, for example Waiting: 2 agents, 1 task, and the sidebar keeps the session’s dot busy, so a chat never looks done while work is still running.
Stopping the turn with Escape stops its subagents too.
Reopened sessions
When you reopen a session, its lanes are read back from disk, so you can still read what each subagent did.