A migration has three steps, and the third can only be written once the second has found which call sites actually change. You want all three delegated. What should you ask for, and what will the third worker have to go on?

The principle — dependency decides the arrangement, and the arrangement decides what the next worker knows. Those are two separate questions and this one asks both.

Why the key is correct — Anthropic's instruction for multi-step workflows is to ask Claude to use subagents in sequence: each subagent completes its task and returns results to Claude, which then passes relevant context to the next subagent. Both halves answer the question. The sequence is forced, because step three's input does not exist until step two has run. And the hand-off travels through Claude rather than between the workers, so what the third subagent gets is the relevant context Claude gives it, assembled out of what the second one returned.

Why the others are wrong — a chain does not carry context by itself. Each subagent starts with a fresh, isolated context window and sees neither the conversation nor the previous worker's window, so anything missing from the brief is missing outright. Starting all three together is the wrong arrangement for the reason the question already states: fan-out suits paths that do not depend on each other, and there is nothing to reconcile until the second step finishes. And dependent work is not undelegatable — sequencing is how it gets delegated.

Remember this — chain the dependent steps, and treat every link as a brief that has to be written rather than inherited.

Sources — Anthropic's Claude Code subagents documentation.