Debug-gym Can AI agents lighten developers debugging load

The Growing Role of AI in Coding

With the rise of AI coding tools like GitHub Copilot, developers are increasingly relying on AI to generate code. GitHub CEO Thomas Dohmke predicted that "sooner than later, 80% of the code is going to be written by Copilot." This trend is evident in startups, where 95% of code for a quarter of Y Combinator’s latest batch was written by large language models (LLMs).

However, most developers spend the majority of their time debugging code, not writing it. This raises an important question: Can AI tools also assist in debugging?

Introducing Debug-gym

Microsoft Research has developed debug-gym, an environment that equips AI agents with interactive debugging tools. Unlike traditional AI coding tools, debug-gym allows agents to:

• Set breakpoints
• Navigate codebases
• Print runtime variable values
• Create test functions

Key Features of Debug-gym

• Repository-level information: Agents can navigate and edit files across the entire repository.
• Robust and safe: Code runs in sandboxed Docker containers to prevent harmful actions.
• Extensible: Easily add new tools to the environment.
• Text-based: Compatible with modern LLM-based agents.

Early Results and Future Work

Initial experiments show promising results. While current AI tools struggle with complex debugging tasks, debug-gym-enabled agents show significant improvement. For example, on the SWE-bench Lite benchmark, agents with debugging tools performed better than those without.