The Bitter Lesson of Tool Calling
Across 14 models, typed Python tool stubs usually beat or matched native JSON calls.
The paper compares programmatic tool calling with native JSON tool calling on BFCL v4. In this setup, models call typed Python stubs, with execution and results handled inside one agent turn. Programmatic tool calling matches or exceeds JSON in 11 of 14 models, and in 13 of 14 under parallel fan-out. The GPT-5.6 family shows a 10.6% gain over the JSON baseline, while the approach stays steadier under context rot. Source: ArXiv · AI/CL/LG's note.
The paper compares programmatic tool calling with native JSON tool calling on BFCL v4. In this setup, models call typed Python stubs, with execution and results handled inside one agent turn. Programmatic tool calling matches or exceeds JSON in 11 of 14 models, and in 13 of 14 under parallel fan-out. The GPT-5.6 family shows a 10.6% gain over the JSON baseline, while the approach stays steadier under context rot. Source: ArXiv · AI/CL/LG's note.
score 6