Parallel Fan-Out: Wall Clock vs. Token Cost

You need to review 12 independent microservice repos for a specific deprecated API usage before a breaking upgrade next week. Each review takes a single agent roughly 6 minutes.

1. Compute the wall-clock time for reviewing all 12 sequentially vs. fanning them out to 12 parallel subagents. What's the actual mechanism that makes the parallel version faster, not just "more agents run at once"?
2. Does the parallel version cost less, more, or the same in tokens as the sequential version? Justify your answer precisely.
3. Your teammate suggests fanning out to 100 parallel subagents instead of 12, one per file changed instead of one per repo, arguing "more parallelism is strictly better since wall clock only depends on the slowest one." What's wrong with this reasoning in practice?