The latest release of swe-agent/mini-swe-agent is v2.4.6. The repository materials do not state what changed in this version. The project is an AI software engineering agent that helps with solving GitHub issues and command-line tasks. It is written in Python and released under the MIT license.
According to the README, the agent uses only bash to run commands. It is built on a Python agent class of roughly 100 lines and a linear history, and it executes each command separately via subprocess.run. In addition to the CLI, it offers batch inference, a Trajectory browser for looking through the execution process, and Python bindings. It can run not only in local environments but also in Docker and Podman. The repository description cites a score of more than 74% on SWE-bench verified. This is a figure presented by the project, and it is worth examining together with the fact that this is a tool that performs coding tasks on a small codebase.
However, the provided materials contain neither the changes in v2.4.6 nor the evaluation conditions behind that score. There is therefore not enough ground to judge what improved in this release, or to compare the score directly with results from other environments. What can be confirmed at present is limited to the structure of the agent, the features it provides, and the performance figure stated by the repository. Rather than treating it as an independently verified performance result, it is more appropriate to read it as a claim made by the project.