SWEeper-Bench: Evaluating Agents on Interactive Software Bug Discovery
October 7, 2026
SWEeper-Bench introduces 200 tasks derived from real open-source web application bug fixes to test agents' ability to identify interaction-dependent failures. Unlike previous benchmarks, agents must discover bugs through multi-step software interaction rather than static code analysis.
HOW THIS AFFECTS YOU
●
builderYou can now benchmark how well your coding agents handle complex, runtime-dependent bugs.
●
researcherYou can evaluate agentic reasoning in dynamic, interactive software environments.