Workload-Aware Cache Eviction for Multi-Agent DAGs Cuts Latency 64.7% vs Uncached, 31.1% vs the Next-Best Finite Cache

arXiv 2607.20495·medium signal

Mohamed, Haque, Khan, Sharma, Ge, and Anwar point out that standard eviction policies score cached entries on access history alone, ignoring signals that agentic execution makes available for free: recomputation cost, DAG dependency count, and agent invocation frequency. Combining those three into one scoring function reduces latency by up to 64.7% against an uncached baseline and averages 31.1% over the next-best finite-capacity policy across three multi-agent benchmarks, approaching unbounded-cache performance while matching or beating competitors on accuracy. They also show it composes with plan-level caching and parallel agent execution, since each targets a different bottleneck.