Breaking the Role: At What Point Does Compressing a Shared KV Cache Degrade Agent Capabilities?

Multi-agent LLM pipelines increasingly hand off KV caches, but existing analyses focus on sharing uncompressed caches and grades based on the agent that owns the cache. We present a controlled study of compressing a shared reasoning cache to evaluate whether the receiving agent still performs its role. The pipeline consists of two Qwen3-8B agents; the verifier agent inherits the solver's reasoning trace, across a grid of eviction methods (SnapKV, R-KV, StreamingLLM) $\times$ placement (handoff-time, decode-time) $\times$ quantization (HQQ 2/4/8-bit), with memory reported as measured packed bytes. We measure the verifier's functionality on how often it corrects the solver's answer, how often it breaks a correct answer, and the net value in accuracy when compared to running the solver alone. On MATH-500 Level~5, pruning to 2048 tokens plus 4-bit quantization retains 97\% of uncompressed accuracy at 7.1 $\times$ less memory (0.791 vs. 0.813), with the correction rate intact (0.618 vs. 0.632). However, under heavier compression, the verifier becomes less effective, retaining 55\% of its accuracy and only 33\% of its correction rate at 2-bit quantization, while also breaking correct answers (0.000 to 0.303), and the net-value in accuracy turns negative, meaning the pipeline does worse than the solver alone. On AIME25 (n=30), quantization at 4 bits or higher matches or improves upon the uncompressed cache but every eviction condition falls below it. For harder reasoning traces, lower precision preserves higher accuracy compared to dropping tokens.