SWE-agent

SWE-agent enables a language model of choice, such as GPT-4o or Claude Sonnet 4, to autonomously use tools to fix issues in real GitHub repositories, find cybersecurity vulnerabilities, or perform custom tasks. It is an academic project built and maintained by researchers from Princeton University and Stanford University, aimed at developers and researchers working on automated software engineering.

SWE-agent is configured through a single YAML file that governs the agent's behavior, leaving maximal agency to the language model rather than constraining it with a rigid pipeline. An offensive-cybersecurity mode, EnIGMA, adapts the same agent to solve capture-the-flag challenges, and the project's documentation covers installation, a command-line hello-world example, and benchmarking against SWE-bench.

SWE-agent has topped SWE-bench among open-source projects, with SWE-agent 1.0 and Claude 3.7 reaching the top of the SWE-bench full and verified leaderboards, and SWE-agent-LM-32b reaching the top open-weights score. Its EnIGMA mode has achieved leading results on multiple cybersecurity benchmarks, and the project's own mini-swe-agent variant matches SWE-agent's performance in about 100 lines of Python.

SWE-agent is released under the MIT license and can be tried directly in a browser or installed from the command line. The project's current development effort has moved to mini-SWE-agent, which matches SWE-agent's performance while being much simpler, and the maintainers now generally recommend using mini-SWE-agent instead of SWE-agent going forward; EnIGMA users are advised to stay on SWE-agent 0.7 until it is updated for 1.0.

Use cases

• Automatically fixing a bug reported in a GitHub issue
• Solving a capture-the-flag cybersecurity challenge with EnIGMA
• Competing in a competitive coding challenge