Cache the agent's execution plan, not the token sequence, when user phrasing varies

Harness Engineering Academy·medium signal

Two 2026 techniques extend caching past byte-identical prefixes. Semantic caching (vCache, introduced Feb 2026) embeds the incoming prompt and serves a cached response when similarity clears a threshold — useful when intent is identical but wording differs. Agentic plan caching goes further by caching the execution plans planning agents generate, so repeated task shapes skip re-planning entirely rather than re-deriving the same steps.