Curious what kinds of bugs the multi-agent setup catches thatsingle-pass review misses in practice. Is it more about coverage(different agents looking at different aspects) or about getting a second opinion on the same aspect? The README has examples but the mechanism by which the parallelism actually helps isn't obvious to me from them.