What Limits Agentic Systems Efficiency?

Large Language Models (LLMs), such as OpenAI-o1 and DeepSeek-R1, have demonstrated strong reasoning capabilities. To further enhance this process, recent agentic systems, such as Deep Research, incorporate web interactions into LLM reasoning to mitigate uncertainties and reduce potential errors. However, existing research predominantly focuses on reasoning performance, often neglecting the efficiency of these systems. In this work, we present a comprehensive empirical study that identifies efficiency bottlenecks in web-interactive agentic systems. We decompose end-to-end latency into two primary components: LLM API latency and web environment latency. Our findings show that both components significantly contribute to the overall system latency. To improve latency, we propose SpecCache, a caching framework augmented with speculative execution to reduce web environment overhead. Extensive evaluations on two standard benchmarks demonstrate that our approach reduces web environment overhead by up to $3.51\times$, without compromising agentic system performance.