Prompt-Cooperative KV Cache Reuse for Efficient Multi-Agent LLM Inference

Multi-agent LLM systems suffer from redundant prefill computation as agents repeatedly process overlapping context---shared questions, peer outputs, and instructions. Prior methods address this by correcting KV cache offsets at inference time, requiring online anchor pools or selective recomputation. We propose a simpler alternative: \textit{prompt-cooperative design}, where agent prompts are structured to minimize KV cache divergence across agents, enabling direct cache reuse without correction. Our key insight is that when system prompts differ by only 2--3 tokens (short role names), the resulting KV cache deviation for shared content remains below 2\% in cosine distance across all layers, making uncorrected reuse safe. We implement this in vLLM via a greedy token aligner that identifies reusable spans and skips their forward computation. On three benchmarks (MMLU, GSM8K, HumanEval) with Qwen2.5-7B-Instruct and Qwen2.5-Coder-7B-Instruct, our approach achieves $\sim$50\% prefill token reduction with $<$0.8 percentage point quality loss compared to standard prefix caching---matching the quality preservation of more complex methods like KVComm while requiring no offset correction machinery.