What it does
SWE-agent is an academic project built and maintained by researchers from Princeton University and Stanford University. You pick the language model, for example GPT-4o or Claude Sonnet 4, and the agent uses tools on its own to work through an issue in a real GitHub repository. The README also describes uses for finding cybersecurity vulnerabilities and for custom tasks, plus a separate EnIGMA mode for offensive-security capture-the-flag challenges. Behaviour is governed by a single YAML file, and the documentation covers installation, a command-line hello world and benchmarking on SWE-bench. There is also a way to try it in the browser. The maintainers say most of their current development effort has moved to mini-SWE-agent, which they describe as having superseded SWE-agent while matching its performance with a much simpler design. They recommend mini-SWE-agent going forward. SWE-agent is released under the MIT licence and was presented at NeurIPS 2024.
Capabilities
- Lets a language model of your choice use tools to work on issues in real GitHub repositories
- Governed by a single YAML configuration file
- EnIGMA mode for offensive cybersecurity and capture-the-flag challenges
- Can be pointed at custom tasks beyond issue fixing
- Documentation covers installation, command-line use and SWE-bench benchmarking
When to use it
- Researching how language-model agents handle software engineering tasks
- Trying automatic fixes for GitHub issues with the model of your choice
- Experimenting with agent-driven security and capture-the-flag work
