An AI agent's ability to act correctly depends not only on the quality of the underlying language model but, to a surprising degree, on how the tools available to it are designed and described.
Descriptions that are clear, not ambiguous
A tool with a vague description — "handles customer data" rather than "updates the email address of an existing customer given their ID" — leads the model to misjudge when and how to use it. Precise descriptions, with clear examples of when the tool should and should not be used, cut errors dramatically.
A narrow scope beats a generic tool
A tool with a very specific job is more predictable than a generic one attempting several different things. Splitting capabilities into focused tools, each with a clear responsibility, makes it easier for the model to pick the right one for each situation.
Required parameters and sensible defaults
Setting out clearly which parameters are required and which optional, with reasonable defaults where appropriate, reduces the occasions on which the model has to guess at something missing. A well-designed tool guides the model towards correct use rather than leaving too much room for interpretation.