SWE-agent is an open-source research agent that takes a GitHub issue and automatically tries to fix it with the language model of your choice. Built by the team behind SWE-bench at Princeton, it has more than 20,600 stars on GitHub, was presented at NeurIPS 2024, and remains actively maintained with updates in October 2026 under the MIT license.
SWE-agent pioneered the idea of the agent-computer interface: instead of letting a model flail at a raw shell, it gives the agent a carefully designed set of commands for viewing, searching, and editing code, plus guardrails that catch common mistakes. That design made it one of the strongest open systems on the SWE-bench benchmark and a reference implementation for software-engineering agents. Beyond bug fixing, the team and community also apply it to offensive cybersecurity challenges and competitive programming tasks.
SWE-agent is also valuable as an instrument for understanding agents. Every run records a full trajectory of commands, observations, and edits, so researchers can study where agents succeed, where they loop, and how interface design changes outcomes. The project family has since grown lighter-weight siblings for quick experiments, but the original remains the most complete reference for issue resolution. Its results helped define realistic expectations for autonomous software engineering: not magic, but a system that genuinely closes a meaningful share of real issues when paired with a strong model.
SWE-agent is free and open source under the MIT license. The only cost is the model API usage consumed while the agent works, billed by your own provider.
SWE-agent is for researchers benchmarking software-engineering agents, and for developers who want a proven open-source starting point for building agents that can understand a repository and submit real fixes.