How we improved GPT-4o multi-step function calling success rate by 4x