CTOOLSET: cache the tool call so the agent stops paying for it twice
Agents call the same tools with the same arguments constantly. CTOOLSET and CTOOLGET cache tool results keyed by tool plus exact arguments, so a deterministic call runs once and serves many.
Agentic workflows are full of repeated tool calls: five agents in a fan-out all look up the same schema, the same weather, the same exchange rate, each one a round trip and sometimes a metered API hit. For deterministic tools — same arguments, same result — that repetition is pure waste. CTOOLSET caches the result keyed by the tool name plus its exact arguments.
CTOOLGET checks that cache before the agent makes the call: a hit returns the prior result instantly; a miss runs the tool and CTOOLSET banks it for the next caller. Because the key is the tool and its exact args, there's no fuzzy-match risk — 'weather in Berlin' and 'weather in Munich' are different keys, served correctly.
Five agents asking one question should cost one answer.
The payoff scales with fan-out. A single agent rarely calls the same tool with the same args twice in a row, but a swarm of agents working a shared problem does it constantly — and a multi-step plan that revisits the same lookup across iterations does too. Caching the tool layer turns N identical calls into one.
The bottom line
Semantic caching saves you from re-asking the model; tool caching saves you from re-calling the world. Together they mean an agent system's repeated work — model and tools alike — gets done once.