NeuralStackStanford University researchers published SWE-Agent on July 17 2026 achieving 92 percent resolution rate on the SWE-Bench software engineering benchmark. The agent combines a 70 billion parameter open model fine-tuned on 1.2 million GitHub issues with a novel planning loop. The work was accepted to NeurIPS 2026.
The system resolves real GitHub issues across 12 popular repositories including Django and scikit-learn without human intervention. Average resolution time measured 4.2 minutes per issue in the test set.
Previous state-of-the-art systems topped out at 84 percent using proprietary models. The Stanford team open sourced both the fine-tuned weights and the full agent framework under an Apache 2.0 license.
The breakthrough demonstrates that targeted fine-tuning on issue-resolution data can close the gap between open and closed models on complex engineering tasks.
Why this matters
The result accelerates the viability of fully autonomous software engineering agents for production codebases. Companies may integrate similar agents into CI pipelines within the next year.
The open release also provides a reproducible baseline for academic and industry labs studying agent reliability and safety.
Future work will extend the approach to multi-repository and cross-language codebases.