SemiAnalysis says agentic workloads account for more than 70% of all inference traffic. It describes sessions with tens or hundreds of turns, where context accumulates quickly. Reusing cached context can avoid recomputation, depending on available storage, while short-lived sub-agents create bursts of cache activity. A user sharing the report in September 2026 predicts agents will account for nearly all inference within a year or two, pointing to background tasks such as code review, data processing and recruiting research.