DeepSWE: A contamination-free benchmark for long-horizon coding agents

DeepSWE measures frontier coding agents on original, long-horizon software engineering tasks.