What it is

SWE-agent is a Python, MIT-licensed open-source project in the AI-assisted coding and agent ecosystem. It is an agent-based model for developer tools that takes a GitHub issue and tries to automatically fix it using a chosen language model. The project is associated with NeurIPS 2024 and is maintained by researchers from Princeton University and Stanford University.

The concrete problem it addresses is manual repository maintenance: a maintainer must read an issue, reproduce it, inspect code, edit files, test, and submit a change. SWE-agent packages that workflow into an autonomous loop in which the language model uses tools to work on real GitHub repositories. It also extends the same idea to offensive cybersecurity challenges and custom coding tasks.

Key capabilities

• It takes a GitHub issue as input and attempts to fix the issue in a real repository with a chosen language model.
• It lets the language model autonomously use tools, rather than forcing a fixed step-by-step script, and the README describes this as leaving maximal agency to the model.
• It uses a configurable setup governed by a single YAML file, so users can define or adjust agent behavior without editing multiple configuration sources.
• It includes EnIGMA, a mode for offensive cybersecurity capture-the-flag challenges, with results reported on multiple cybersecurity benchmarks.
• It supports custom tasks, coding challenges, and SWE-bench benchmarking workflows, including batch mode.

Who uses it and how

• Researchers and developers run the command-line hello-world workflow to fix a GitHub issue with a chosen language model.
• Benchmarking users run batch mode to evaluate SWE-agent on SWE-bench rather than on one repository issue.
• Cybersecurity participants use EnIGMA for capture-the-flag tasks, while the README says to use SWE-agent 0.7 until EnIGMA is updated for 1.0.
• Contributors use issues and pull requests for changes, with larger code changes discussed in issues first.

Getting started

The README recommends trying SWE-agent through GitHub Codespaces and also directs users to source installation documentation. Common workflows are the command-line hello-world run, SWE-bench batch mode, and EnIGMA cybersecurity mode, with EnIGMA currently pointing to SWE-agent 0.7.

When to use it — and when not to

Use SWE-agent when the full SWE-agent research framework, SWE-bench benchmarking, or EnIGMA cybersecurity workflows are needed. Do not choose it as the default new project if the simpler mini-SWE-agent is acceptable, because the README says most current development effort is on mini-SWE-agent and recommends it going forward. Also, EnIGMA is not fully aligned with the 1.0 codebase, since the README instructs users to use SWE-agent 0.7 while EnIGMA is updated for 1.0.

@inproceedings{yang2024sweagent,
  title={{SWE}-agent: Agent-Computer Interfaces Enable Automated Software Engineering},
  author={John Yang and Carlos E Jimenez and Alexander Wettig and Kilian Lieret and Shunyu Yao and Karthik R Narasimhan and Ofir Press},
  booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
  year={2024},
  url={https://arxiv.org/abs/2405.15793}
}

@misc{abramovich2024enigmaenhancedinteractivegenerative,
      title={EnIGMA: Enhanced Interactive Generative Model Agent for CTF Challenges},
      author={Talor Abramovich and Meet Udeshi and Minghao Shao and Kilian Lieret and Haoran Xi and Kimberly Milner and Sofija Jancheska and John Yang and Carlos E. Jimenez and Farshad Khorrami and Prashanth Krishnamurthy and Brendan Dolan-Gavitt and Muhammad Shafique and Karthik Narasimhan and Ramesh Karri and Ofir Press},
      year={2024},
      eprint={2409.16165},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2409.16165},
}