"Our key insight is that LMs require carefully designed agent-computer interfaces (similar to how humans like good UI design). E.g. When the LM messes up indentation, our editor prevents it and gives
FreeOpen-source AI agent that autonomously fixes GitHub issues
About "Our key insight is that LMs require carefully designed agent-computer interfaces (similar to how humans like good UI design). E.g. When the LM messes up indentation, our editor prevents it and gives
SWE-agent is an open-source system developed by Princeton NLP that autonomously resolves issues in GitHub repositories. It uses a carefully designed agent-computer interface—a specialized terminal—that allows GPT-4 to efficiently edit and run code. The system achieves accuracy comparable to Devin on the SWE-bench benchmark, averaging just 93 seconds per task. Key capabilities include opening, scrolling, and searching files; editing specific lines with automatic syntax checking; and writing and executing tests. The custom interface is critical for performance, preventing common LM errors (e.g., indentation mistakes). SWE-agent is easily configurable and extendable, making it a valuable tool for both software maintenance and research on software engineering agents.
Key Features
Pros & Cons
- Open source with active community contributions
- High accuracy on SWE-bench, comparable to state-of-the-art systems
- Fast average resolution time (93 seconds)
- Custom interface reduces common LM errors like indentation mistakes
- Easily configurable and extendable for further research
- Requires GPT-4, a proprietary and costly model
- Limited to code editing tasks within GitHub repositories
- May not handle complex software engineering tasks beyond code changes