If you have a clear performance metric to optimize, agent swarms can do a great job.

This is what Shopify's Tobi found.

Interestingly, in his case, there weren't any magic bullets.

It was just an accumulation of tons of small benefits.

Humans wouldn't be patient enough to chase these small wins that accumulate.

But LLMs have infinite patience, and can also be parallelized with basically zero cost.