Tool-Use Evaluation
Tool-use evaluation checks whether an AI agent picks the right external tool (a search API, a calculator, a database query) for a given task, calls it with correctly formatted arguments, and correctly interprets the result — independent of whether its final natural-language answer sounds right.
A model can produce a perfectly fluent, confident-sounding answer while having called the wrong tool, passed it malformed arguments, or ignored what the tool actually returned — the fluency of the output gives no signal about whether the underlying tool call was even valid, which is why this needs its own dedicated check rather than being inferred from answer quality.
In practice this means asserting on the tool call itself: was the right function selected, do the arguments match the tool's schema (right types, required fields present, no hallucinated parameters), and does the agent's final answer actually reflect what the tool returned rather than the model's own prior belief overriding it.
A common failure mode worth a dedicated test: the agent calls the right tool, gets a valid result back, and then contradicts it in the final response anyway — the tool call passes, but the feature is still broken from the user's perspective.
