Built this mostly to see if parallel workers would actually beat one strong Codex agent once tasks got bigger.

So far they haven’t. On the largest benchmark Sol had 6 independent modules it could delegate, but chose 0 workers. Forcing 6 workers was around 2x slower.

Repo has the raw results + benchmark setup if anyone wants to poke holes in it. I’d be curious if people see a different result on much bigger real-world repos.