Which format is good for AI Agents? published my benchmarks to Github https://github.com/SamuelLHuber/llm-agent-format-benchmarks