In the programming community, a brand-new open-source project is causing a stir: mini-SWE-agent, developed by the team behind SWE-bench and SWE-agent. The project's highlight is that it can fix 65% of bugs in real GitHub projects with just 100 lines of code, matching the performance of the original version, but with an extremely lightweight architecture.
The design philosophy of mini-SWE-agent is to simplify as much as possible, making it easier for more developers to get started. Compared to the previous SWE-agent, this mini version removes complex tool call interfaces, executing commands only through a basic Bash environment. Each step outputs a complete Shell command from the language model, significantly improving compatibility without relying on multiple plugins or tools.
This project not only reduces dependencies, but also compresses the code of mini-SWE-agent to about 100 lines, and with the required environment and model, the total code is less than 200 lines. This minimal architecture allows developers to focus more on coding itself, rather than being troubled by complicated configurations and tools.
mini-SWE-agent retains powerful features, still solving about 65% of problems in the SWE-bench verification. It also comes with tools such as batch inference and trajectory browser, helping users perform large-scale evaluation and decision analysis. In addition, developers can quickly start the agent via a command-line tool, and even provide a visual interface, making it convenient for developers to monitor the execution process in real time.
The team behind the project stated that mini-SWE-agent is more suitable for developers who want to run quickly and pursue a simple control flow, while SWE-agent is more suitable for users who need high configurability and complex history state management. This flexible choice makes mini-SWE-agent an attractive tool for daily development.
mini-SWE-agent embodies the development concept of strong readability and easy extensibility. It can be used as a command-line tool, and also integrated into other Python applications, helping developers solve problems efficiently.
Project address: https://github.com/SWE-agent/mini-swe-agent