SEMAG: Self-Evolutionary Multi-Agent Code Generation

A research team led by Yulin Peng and Haowen Hou has published a paper on arXiv introducing SEMAG (Self-Evolutionary Multi-Agent Code Generation), a novel framework designed to overcome limitations in current AI coding assistants. Unlike static systems that rely on manual model selection and fixed workflows, SEMAG dynamically decomposes programming tasks into stages—planning, coding, debugging, and discussion—and adapts its process based on task complexity. Its key innovation is a team of "self-evolutionary" AI agents that can access and automatically upgrade to the latest, most suitable large language model (LLM) in real-time, ensuring the system continuously improves as new models are released.

SEMAG has set new state-of-the-art benchmarks for code generation. When using the same backbone LLMs as previous methods, it achieved a 3.3% higher Pass@1 accuracy on the challenging CodeContests benchmark. More impressively, when its self-evolutionary model selection feature is activated—allowing it to automatically identify and switch to the optimal AI model for a given task—its performance jumps to 52.6%. This demonstrates the framework's dual strength: an effective multi-agent architecture for complex problem-solving and built-in adaptability to the rapidly evolving landscape of foundation models. The approach marks a significant step toward more autonomous and capable AI software engineering tools.