Skip to main content
The agent can spawn a subagent — a separate background agent loop that works a delegated task independently, with its own model and its own tool catalog, while you keep chatting in the main session. The Subagents rail tab, listing subagents with their status and a detail view with a transcript

​
Spawning one

A subagent is created by the spawn_subagent tool, which the agent calls either because you asked it to delegate something, or on its own judgment for a self-contained chunk of work. Either way, it’s required to know which model to run the subagent on before it spawns it:
1

Resolve the model

If you already named a model (or a clear preference like “the cheap one”), the agent resolves that itself by calling list_models — which returns every model’s real cost in credits, paid/free status, and capabilities.
2

Ask if it's not specified

Otherwise, it calls ask_user_question to ask you directly, with concrete one-click options built from real model data — never a silently-guessed model.An ask_user_question card with one-click model options and a free-text fallback
3

Spawn

spawn_subagent launches the background loop and returns immediately. You’ll see a short confirmation in the transcript; the subagent keeps running while you do other things.
Like any mutating tool, spawn_subagent itself goes through the normal permission gate — Ask by default, bypassed in Auto Mode.

​
What a subagent can do

A subagent runs unattended — nobody is watching its loop turn by turn the way you watch the main chat. What it’s allowed to do depends on the main session’s mode at the moment it’s spawned.
A subagent can’t spawn its own subagents, and can’t call ask_user_question — there’s nobody there to answer it mid-flight. It otherwise has the same tools as the main agent: reading and editing scripts, properties, instances, and (per the table above) execute_luau.

​
Subagent Views

Changed in v3. Subagent Views moved out of a pill-triggered overlay and into its own tab in the left icon rail. The rail icon itself now carries a small unread dot the instant a subagent finishes — paired with the same toast as before — so you find out without having to go looking, whichever tab you’re on. A small gray line also drops straight into the main chat transcript itself (“Subagent done: label”), so you don’t even have to switch tabs to notice.
Opening the tab shows every subagent, newest first, with a status pill (Running / Done / Error / Stopped) and clears the unread dot. Selecting one opens its full transcript, plus:
  • A Stop button while it’s running.
  • Any pending permission request, if the main session wasn’t in Auto when it was spawned — resolved here, exactly like a normal tool-call card.
  • A follow-up box once it’s finished, to send it another instruction and resume it in the background.

​
Checking on a subagent without flooding context

The main agent has two tools for this, used instead of waiting idle:
​
read-only
Returns status, label, model, and the subagent’s most recent message only — never its full transcript, so checking in doesn’t blow up the orchestrator’s own context.
​
mutating
Sends a follow-up instruction and resumes a subagent, whether it’s still running or already finished.
When a subagent finishes, its completion is queued and flushed into the main session’s context the next time you send a message — the orchestrator finds out without the plugin spending credits on an extra, unprompted model call just to relay the news.
Subagents are in-memory only for this release — they don’t persist across a Studio reload. Closing and reopening Studio clears Subagent Views.