A new AI system automatically finds and fixes bugs in code written by other AIs.

Researchers have developed TraceCoder, a multi-agent AI framework that automatically debugs flawed code generated by large language models. It works by inserting probes to capture detailed runtime data, analyzing the root cause of failures, and learning from past mistakes to avoid repeating them. In tests, it improved code accuracy by up to 34.4% over leading methods, with its iterative repair process alone contributing a 65.6% gain.