SWE-agent came out of a research collaboration between Princeton and Stanford focused on a specific question: how much of an autonomous coding agent's performance comes from the underlying model versus the interface it's given to interact with a codebase. Its answer was to build a custom agent-computer interface, a restricted but carefully designed set of commands for viewing, editing, and searching code, rather than giving the model raw shell access.
That interface design turned out to matter a lot. SWE-agent posted some of the strongest results on SWE-bench, the standard benchmark for resolving real GitHub issues, when it was released, and its approach influenced how several later commercial coding agents thought about constraining what an agent can do at each step rather than giving it unrestricted freedom.
It's free, open source, and primarily a research artifact rather than a polished product: useful for understanding how agent-computer interface design affects reliability, and as a reference implementation for teams building their own coding agents rather than something most developers would install for daily work.