Delegation is allowed. When work splits into genuinely independent parts — different files, no shared state — running them in parallel is fine.
What still holds:
- The primary agent owns the result. A subagent's report is a claim, not a verification: check the tests and the code yourself before believing it.
- Parts that depend on each other are not independent. Two agents editing one file will clobber each other's work regardless of how the task was described.
- Every part still has to be verified the same way: tests run, lints clean, and a failure reported as a failure.