CodeAgents + Structure: A Better Way to Execute Actions

Hugging Face · 1 year ago · 18

Researchers combined code-based actions with structured JSON generation to improve AI agent performance, requiring agents to output both explicit reasoning and executable Python code in a constrained format. Across benchmarks including GAIA and MATH, structured CodeAgents outperformed traditional CodeAgents by 2–7 percentage points on capable models, while parsing errors in unstructured code reduced success rates by 21.3 percentage points. The approach works best with large, well-trained models (32B+ parameters) but creates a "structure tax" for smaller models that struggle with simultaneous JSON formatting, Python syntax, and problem-solving demands.