You have spent an hour with Claude settling a migration design. You want a second opinion from a worker that already knows the whole discussion, and you then want that worker to pass the same whole discussion on to several more workers of its own. What can you actually get?

The principle — an inherited conversation is not a transferable asset. It arrives once, at the moment the worker starts, and it goes no further.

Why the key is correct — the two halves of the plan are settled by two different sentences. A fork inherits everything the main session has at the moment it spawns, while any other subagent starts fresh from its definition, so the second opinion from a worker that already knows the hour of discussion is available. But a fork cannot spawn further forks, so passing that same whole discussion on to several more workers is not, and the plan gets its first half only.

Why the others are wrong — the inheritance cannot be requested; it is a property of the fork rather than an instruction you give a worker. Ruling out both halves gives up the part that works, since starting from a definition is true of every subagent except this one. And inheriting a conversation does not carry the right to hand it on, which is the limit the second half of the plan runs into.

Remember this — a fork is one level deep by design. If work needs splitting further, brief the next workers yourself.

Sources — Anthropic's Claude Code subagents documentation.