Mini Programming Agent Fixes 65% of Bugs in Just 100 Lines
Mini Programming Agent Achieves High Bug-Fix Rate with Minimal Code
In a significant breakthrough for the programming community, the mini-SWE-agent project has demonstrated remarkable efficiency by resolving 65% of real-world GitHub project bugs using only 100 lines of core code. Developed by the team behind SWE-bench and SWE-agent, this open-source tool offers a streamlined alternative to traditional debugging methods.
Simplified Architecture for Broader Accessibility
The mini-SWE-agent was designed with simplicity as its guiding principle. Unlike its predecessor, this version eliminates complex tool-call interfaces, relying instead on a basic Bash environment for command execution. Each step generates a complete Shell command directly from the language model, enhancing compatibility without requiring multiple plugins or additional tools.
Lightweight Yet Powerful Performance
With its minimalist architecture, the entire package—including environment setup and model—comprises fewer than 200 lines of code. This approach allows developers to focus on coding rather than configuration. Despite its reduced size, mini-SWE-agent maintains impressive functionality, solving approximately 65% of issues in the SWE-bench verification tests.
The tool also includes practical features such as:
- Batch inference for large-scale evaluation
- Trajectory browser for decision analysis
- Command-line interface for quick deployment
- Visual monitoring of execution processes
Flexible Implementation Options
The development team emphasizes that mini-SWE-agent caters to programmers seeking rapid execution and straightforward control flows. For users requiring more complex state management or higher configurability, the original SWE-agent remains available. This flexibility makes mini-SWE-agent particularly attractive for daily development tasks.
The project exemplifies principles of readability and extensibility, functioning both as a standalone command-line tool and as an integrable component within larger Python applications.
Key Points:
- Resolves 65% of real GitHub project bugs with minimal code
- Operates through basic Bash environment without complex dependencies
- Entire package under 200 lines including environment setup
- Maintains original SWE-bench performance despite simplified architecture
- Offers both command-line and visual interface options
The mini-SWE-agent is available at: GitHub Repository
