Evolver lets coding agents reuse past experience, raising task success in tests

Researchers introduce Evolver, a framework that stores agent exploration as reusable Gene assets, according to an X post by Elvis Saravia.

On a 320-run diagnostic set across Terminal-Bench 2.1, SWE-bench Multilingual, SWE-rebench and GAIA, task success rose from 83.13% to 87.50%, and recorded token use fell 24.24%. The post says the gains came without changing the underlying models. Evolver is described as working with Claude Code and Codex. The source is a single social media post, and the results have not been independently verified.