📝Prompt Caching is an LLM API feature that stores a stable prefix of a request, such as system instructions and tool definitions, so later calls reuse it instead of reprocessing it, cutting both cost and latency.
Agent requests repeat most of their content on every turn: the system prompt, tool schemas, project instructions, and earlier conversation. With caching, the provider saves the processed state of that prefix after the first call. Later requests that begin with the identical prefix read from the cache, and only the new tokens at the end are processed fresh.
The savings are largest on long, repetitive prompts. 📝Anthropic bills cache reads at a fraction of the base input price and keeps a cached prefix for five minutes by default, with a one-hour option at extra cost; other providers offer similar mechanisms. Any change early in a prompt invalidates everything after it, so cache-friendly prompts put stable content first and variable content last.
Caching is one of the main levers in 📝Token Optimization. In 📝Claude Code, cache lifetime is controlled through environment variables such as 📝ENABLE_PROMPT_CACHING_1H, and agent platforms such as OpenClaw build their own caching layers on the same principle.