Overview
SWE-agent is an academic project (Princeton / Stanford et al.) providing an automated software engineering agent framework. It focuses on reproducible, configurable agent-driven workflows and offers benchmarks (SWE-bench), examples, and full documentation.
Key features
- Multiple examples and modes (including SWE-bench and Mini-SWE-Agent).
- Comprehensive documentation and quick-start guides; try demos in GitHub Codespaces.
- Research-oriented: configurable, reproducible, and benchmark-ready.
Use cases
- Research prototypes for automated code repair and repository maintenance.
- Security research and CTF-style evaluations (EnIGMA mode).
- Benchmarking LLMs on software-engineering tasks using SWE-bench.
Technical details
- Implemented in Python with a documented website at https://swe-agent.com .
- MIT licensed; repository includes CI, docs build, and tests.
- Provides workflows and interfaces for converting tools/functions into executable units and evaluation harnesses.