Real-SWE benchmark tests coding agents on private codebases from real companies, with error bars

daveholtz · x · 2026-09-11

janaksunil's team launched Real-SWE, a coding benchmark built on private, out-of-distribution codebases. Agents are evaluated on real software engineering tasks performed by engineers at actual companies — an app with 200K+ users, a fintech platform processing 100K+ bank statements, enterprise sales tools — with reportedly surprising results. daveholtz highlights that it uniquely includes error bars.

Original post →

More from coding & agent

coding & agent channel →