TraceCoder: A Trace-Driven Multi-Agent Framework for Automated Debugging of LLM-Generated Code

Researchers specializing in artificial intelligence published a technical paper on the arXiv preprint server on February 11, 2025, introducing TraceCoder, a novel multi-agent framework designed to fix bugs in code generated by Large Language Models (LLMs) more effectively. The team developed this system to address the persistent issue of subtle but critical errors in AI-generated code, which traditional repair methods often fail to catch because they rely on simple pass/fail signals rather than deep behavioral analysis. By utilizing trace-driven data, the framework aims to provide a more granular understanding of how programs execute and why they eventually fail during complex tasks. The core innovation of TraceCoder lies in its move away from superficial debugging. Standard automated repair tools typically operate in a black-box fashion, identifying that a code snippet is broken without understanding the logical path that led to the crash. TraceCoder disrupts this cycle by employing multiple specialized AI agents that analyze the execution trace—a step-by-step record of the program's internal state. This visibility allows the system to pinpoint the exact location of a logic error rather than guessing based on output alone, representing a significant shift toward 'white-box' automated debugging. Beyond just identifying errors, the TraceCoder framework is built to solve the problem of repetitive failure cycles. Many current LLM-based repair processes tend to hallucinate similar incorrect solutions multiple times because they lack a memory of previous unsuccessful attempts. TraceCoder incorporates a learning mechanism that allows it to refine its troubleshooting strategy based on prior failures. This iterative intelligence ensures that the debugging process becomes increasingly efficient, reducing the computational overhead and time required to produce functional, production-ready code.