The value you get out of a simpler adversarial loop to critique your "main" agent's work is high. Stacking Steve Yegge's personal Kingdom of Nouns on top of each other doesn't add much more.

And this doesn't even begin to get into the madness that is verification for software that matters and is exposed through multiple modalities. You cannot let an agent just vibe its way around "does this business-critical thing with these specific use cases do its job correctly", much as Yegge might have you believe.