The Bitter Lesson of Tool Calling

The paper, titled 'The Bitter Lesson of Tool Calling', evaluates whether replacing rigid JSON tool calls with script-based calls enables more natural chaining and parallelization for LLM agents. It is the first systematic comparison of 'tools as code' across current and prior model generations on a standard benchmark. The title references Rich Sutton's 'bitter lesson' argument, suggesting that scalable, general methods (like code) outperform hand-crafted formats. The findings likely have practical implications for designing tool-use interfaces in production LLM systems. The abstract indicates the evaluation uses real-world task conditions, making the results more actionable for developers.