Scaling Auto-Research Loops for Efficient Agent Harnesses

I found this work very interesting, specially the section about having a swarm (or 5 small teams.) and make the savings really matter.

Not sure how useful it may be to me, in my daily use. My tasks are way more encapsulated (e.g. edit the CSS of this website for me) than the "long horizon" that Edge-Bench aims to represent (e.g. working on a math theorem)