SWE-agent is an academic framework from Princeton and Stanford researchers that lets a language model of your choice autonomously use tools to fix issues in real GitHub repositories, tackle coding challenges, or solve offensive cybersecurity (capture-the-flag) problems via its EnIGMA mode. Its research contribution is the "agent-computer interface" idea — giving the LM a carefully designed set of tools while leaving it maximal agency — and it reported state-of-the-art results on SWE-bench among open-source systems. The whole agent is governed by a single YAML configuration file and is deliberately simple and hackable, aimed at researchers as much as practitioners.
The maintainers now direct most development effort to a successor, mini-swe-agent, which they say matches SWE-agent's performance while being much simpler (the README cites 65% on SWE-bench Verified in about 100 lines of Python), and they recommend it over SWE-agent going forward. SWE-agent sits in a broader family of related projects from the same group, including SWE-bench (the benchmark), SWE-ReX, and SWE-smith; the EnIGMA cybersecurity mode currently requires the older 0.7 release while being updated for 1.0.