Weco preprint: agent self-improves its own software 7 times over 8 days, gains transfer to 4 benchmarks

VraserX · x · 2026-09-25

VraserX cites Weco's new preprint to push back on the lazy "it's just a wrapper" framing: the study let an agent autonomously make seven improvements to its own software during an eight-day run, and the gains transferred to four held-out benchmarks — with the underlying models held fixed. The takeaway: improving how an agent searches, manages context and checks its own work can matter a lot on its own, independent of model upgrades.

Original post →

More from coding & agent

coding & agent channel →