An Empirical Boundary Analysis of Vector-Quantized Semantic Transport

Multi-agent language-model systems move most of their internal traffic — routing labels, handoffs, tool selections — as natural language, an expensive detour through a human-readable format no human reads. A recent line of work removes this cost by having models communicate in continuous latent space (KV-caches or hidden states), but a continuous representation cannot be consumed by the text APIs deployed models expose, and it is not hashable, so it cannot be deduplicated, cached by exact match, or version-tracked. Vector-Quantized Semantic Transport (VQT) keeps the inter-module message discrete: a single integer index into a learned codebook. Discreteness makes the message a hashable key — cacheable, dedupable, and version-trackable — while a small one-label-per-code dictionary keeps the channel compatible with any text-API model on a cache miss. No language-model fine-tuning is required; the only training is a k-means codebook. Rather than reporting reconstruction fidelity, we measure the quantity that matters for a control plane: how much routing information survives quantization. On intent-classification benchmarks with control-plane structure (CLINC150, Banking77), an 8-bit code retains 91–96% of full-precision routing accuracy at 1536× representational compression, with a streaming deduplication rate of 0.89–0.92. End-to-end with gpt-4o-mini on the Berkeley Function-Calling Leaderboard, the channel preserves routing to within 1.8 points of the raw-message baseline (0.912 vs. 0.930). Codebook mismatch silently collapses routing to below 2%; a lightweight header version identifier restores it completely. The scope is deliberately narrow: a fast, deterministic, text-compatible primitive for the control plane of systems that already compute sentence embeddings — not a replacement for natural language in precise, data-plane transfer. Code, experiment harness, and results accompany the paper.