PITH: Inter-Agent Payload Compression Using Zipf-Based Density Scoring and Benford's Law Structural Integrity Validation

This preprint introduces PITH, a lightweight compression framework for inter-agent payloads in multi-agent LLM systems. Unlike existing tools that compress the initial prompt (LLMLingua) or the final user-facing output (Caveman), PITH addresses the agent-to-agent communication layer tool results, reasoning traces, and intermediate context an underexplored source of token waste in complex pipelines. The central contribution is the application of Benford's Law as a post-compression structural integrity validator. We demonstrate empirically that sentence-length distributions in human-written text approximate the Benford distribution (mean MAD = 4.07%), while AI-generated text shows systematic deviation (mean MAD = 10.57%, a 2.6× gap). This signal serves as a feedback gate: if compression increases deviation beyond a threshold, the system relaxes the ratio and retries automatically. Sentence selection uses a Zipf-based density scorer (word length as a lexical rarity proxy), requiring no external dependencies or API calls. PITH achieves 30–60% token reduction on typical agent outputs while preserving code blocks, JSON, URLs, and numerical data verbatim.