Verified 2026-08-05 What was checked, and when — not a guarantee it is still true.

What changes when my skill dispatches sub-agents?

Sub-agents run under the same containment as the main thread — they are not a way to escape a path restriction or a permission prompt. A finished sub-agent cannot be resumed in Cowork, so recovery means dispatching a fresh one with the context restated. If you dispatch without naming an agent type you get a general-purpose agent holding the full tool surface, which is rarely what a skill intends.

Always name the agent type when dispatching a sub-agent.

Measured

Omitting the type yields a general-purpose agent holding the full tool surface, including the shell. If your intent was a constrained helper — a reader, a summariser — you have instead created something with more capability than the task needs, and the dispatch will look correct in every log.

Caveat: Observed in a large sample of real dispatches, where the type-less form was a substantial fraction of the total.

Do not plan to resume a finished sub-agent. Dispatch a fresh one with the context restated.

From binary

In Cowork a completed or stopped sub-agent cannot be continued — the mechanism that does this elsewhere is not available. Recovery and follow-up work therefore mean a new dispatch carrying whatever context it needs, so design sub-agent prompts to be self-contained.

Do not use a sub-agent to get around a path restriction or a permission prompt.

Measured

Sub-agents run under the same containment as the main thread. The path rules apply identically inside them, and the permission machinery fires there too. Dispatching a sub-agent changes what work happens in parallel; it does not change what is allowed.

Put everything a sub-agent needs in its prompt. It cannot ask you for more.

Inference

Because a finished sub-agent cannot be resumed, a dispatch that omits context cannot be topped up — it can only be redone. Restating context is cheaper than a failed dispatch, and it is the only recovery mechanism available.

Caveat: Follows from the no-resume behaviour rather than being separately measured.

Write sub-agent fan-out so that a dispatch being queued or refused is handled, not fatal.

From binary

Concurrency and per-session limits are enforced, and the numbers have changed between releases. A skill that dispatches a fixed wide fan-out and assumes every one starts immediately will behave differently across versions. Treat parallelism as an optimisation, not a requirement.

Write sub-agent prompts with relative paths, exactly as you would the main thread.

Measured

A sub-agent starts in the session's own working area, not in a private one. The same path rules apply, and absolute session paths handed to file tools are rejected there for the same reason.

What is not established

What the confidence labels mean
Measured
Observed live in a real session, with a control where noted.
From binary
Read out of a shipped artifact; the behaviour was not exercised.
Inference
Stated as inference in the source material — read the caveats.