Media report · AI-ASSISTED BRIEF
Google research: self-improving agent harness lifts SWE-bench Lite success from 5% to 40%
Report publication · YYSI publication
According to IT Home, Google researchers published a paper on self-improving agent harnesses, the scaffolding around a model that manages context, tools and verification. Letting an optimizer model edit one of four harness modules at a time, keeping only changes that pass validation, raised SWE-bench Lite success from 5.0% to 40.0% with Gemini 3.8 Flash, at about 46,700 tokens per task. Unrestricted rewrites reached 42.5% but cost nearly ten times the tokens.
Source & context
www.ithome.com · Read the original source ↗
This AI-assisted brief attributes information to the linked media report. It is not an official announcement or independent YYSI reporting. Consult the original report and its evidence.