Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse natural language processing and other complex tasks. However, multi-agent deployments of LLMs often suffer from significant inference overhead caused by a great deal of context and repeated tokens across agents, resulting in high Time-To-First-Token (TTFT) and elevated computational costs. To address these challenges, we propose EACC (Efficient Agent Context Cache), a system that enables position-independent reuse of key-value (KV) caches across agents by semantically chunking the prompt and managing reusable KV pairs through a globally accessible Agent Context Cache. EACC decouples context storage from computation, supports distributed deployment, and integrates with existing Position-Independent Caching (PIC) techniques such as CacheBlend and EPIC for selective recomputation. Our evaluation across multiple datasets and LLM backends demonstrates that EACC achieves up to 82.4% reduction in TTFT, while maintaining output accuracy, thereby offering a scalable and efficient inference infrastructure for multi-agent services.