mini-swe-agent

Why we included this project

Developers curious about how modern coding agents actually work will find this unusually easy to trace: the core agent class is about a hundred lines of Python, and you can follow along as it plans, runs shell commands, inspects output, and retries until the issue is fixed. That transparency does not come at the cost of capability, since the same minimal design clears more than 74% of SWE-bench verified tasks, making it a credible lightweight option for teams that want a local command-line agent without adopting a heavy framework. It runs anywhere a bash shell or container runtime is available, and the code later grew into the agent behind Ramp's SWE-bench work, a sign that the minimal approach generalizes beyond a demo. The docs and FAQ are practical for comparing it with the full swe-agent on the same benchmarks, which helps when deciding which one.