Finally, AutoBe achived 100% compilation success rate on small sized model "qwen3.5-27b". The more important thing is, difference of AI model does not affect to result quality. This is the function calling harness. github.com/wrtnlabs/autobe