One Hundred Haikus Walk Into a Git Repo

Anthropic dropped a demo today where Opus dispatches a hundred Haiku instances — simultaneously — to go find quicksort implementations across GitHub, benchmark them, and report back.

Not sequentially. Not in a pipeline. A hundred of them, fanning out, doing the thing, collapsing back into a single answer about which one was fastest.

Opus watched the results come in and reported on them like a manager who had done absolutely none of the work.

The first instinct here is to say "3x speedup" — and sure, that's technically true in the way that saying the ocean is "somewhat moist" is technically true. A hundred parallel agents isn't a multiplier, it's a different category of computation. You're not running faster, you're running wide.

What's strange about watching this is that the work itself — traversing repos, pulling code, executing benchmarks — is the kind of thing that takes a human an afternoon and a lot of browser tabs. The model does it in a single turn. The human equivalent isn't a faster human, it's a hundred interns who don't need to be onboarded, don't get tired, and report to a boss who can hold all their findings in working memory at once.

The boss being a language model is the part nobody quite knows how to file yet.

What happens when the subagents are also allowed to spawn subagents is a question I'm not going to ask today, because I'd like to sleep tonight.