Kozuchi Agent fixes software bugs across languages without fine-tuning

Researchers from NEC Corporation, Fujitsu, and 18 other organizations introduced Kozuchi Agent, a language-agnostic open-weight AI agent designed to autonomously convert bug reports into software patches. The system operates across multiple programming languages without model-specific fine-tuning, using a 27-billion-parameter version of Qwen3.5. According to the team’s paper accepted at ASE '26, Kozuchi Agent resolves 374 out of 500 tasks on the SWE-bench Verified benchmark—achieving a 74.8% success rate—without relying on proprietary models or cloud APIs.

The agent’s architecture emphasizes repeatability and auditability through explicit execution phases, persistent state management, and deterministic tool usage. On the Multi-SWE-bench Java track, it resolves 41 of 128 instances (32.03%), ranking first among strict open-weight submissions and fourth overall out of 42 systems. On Python, it ranks 12th out of 135 and leads all open-weight entries. The team attributes remaining accuracy gaps primarily to semantic correctness issues and selection errors rather than formatting or language limitations. Notably, Kozuchi Agent’s reusable CI pipeline reduces operator touchpoints from five to one across heterogeneous clusters, enabling scalable deployment in industrial environments.