Definitely not saying multi-agents is all you need for SWE-bench haha. I touch on this at the end of the blog post, where I mention jumps in progress require better base models or tooling.