Notation Matters: A Benchmark Study of Token-Optimized Formats in Agentic AI Systems

The study evaluates two token-optimized formats, TOON and TRON, as alternatives to JSON for structured data exchange in agentic AI systems. The benchmarks reveal that TRON achieves up to a 27% reduction in token usage with a 14 percentage point accuracy drop compared to JSON, while TOON offers an 18% reduction with a 9 percentage point accuracy cost, but suffers from cascading parsing failures in multi-turn scenarios. This research is significant for practitioners as it highlights the trade-offs in token efficiency and accuracy when integrating these formats into LLM-based systems, potentially improving the performance of agentic applications.