Member-only story
CCA-F Part 3: A tool that “worked” can still sink the whole agent. This small, low-scoring exam domain is testing whether you know the difference, and most prepared candidates do not. #
A tool that swallows a timeout and returns an empty list does not just fail, it teaches your model a falsehood and watches it act on it. Return a structured error the model can reason about, and the same failure becomes a recovery instead of a confident lie to your user.
In this article:You will learn why a tool that returns a success flag on a failed call quietly destroys agent reliability, and how to design tools that fail legibly instead. We cover the difference between an unreachable source and an empty result, why tool descriptions are the real router, how
tool_choicesets the model’s latitude, why too many tools is its own failure, and where MCP configuration and secrets belong. The thread tying it all together is one posture: fix the tool, not the prompt.