SWE-agent lets a language model of your choice, such as GPT-4o or Claude Sonnet 4, use tools autonomously to fix issues in real GitHub repositories, solve capture-the-flag cybersecurity challenges, or carry out custom tasks. Its design leaves maximal agency to the model, and a whole run is governed by a single YAML configuration file.
Development effort has moved to mini-swe-agent, which the project describes as matching SWE-agent's performance with a much smaller codebase (65% on SWE-bench Verified in about 100 lines of Python); the general recommendation is to use mini-swe-agent going forward. SWE-agent 1.0 with Claude 3.7 was reported as the top result on SWE-bench Verified and Full at release, and the SWE-agent-LM-32b model reached the top open-weights score.
EnIGMA is a mode of SWE-agent for offensive cybersecurity (capture-the-flag) challenges, with results on several cybersecurity benchmarks; it currently requires SWE-agent 0.7 while it is updated for 1.0. Related projects in the same family include SWE-ReX, SWE-bench, SWE-smith, and sb-cli.
Features
- GitHub issue fixing: point the agent at a repository issue and it produces a patch
- Model choice: any supported language model via API
- Single YAML config: agent behaviour, tools, and environment are defined in one file
- Batch mode: run and score on SWE-bench
- EnIGMA: offensive cybersecurity mode with summariser and interactive commands
- Research orientation: an academic project from Princeton and Stanford, described in a NeurIPS 2024 paper
- Codespaces: try it in the browser through GitHub Codespaces
