Designing a Tool the Model Can Use

Anthropic built a coding agent for a well-known software engineering benchmark, and it reports the following about that work: "we actually spent more time optimizing our tools than the overall prompt."

That surprises many engineers, who expect the instructions to hold an agent's intelligence. Yet the model chooses among tools on every step, from nothing except what the tool definitions say. A definition must therefore serve a reader that cannot open your code, ask a colleague, or try the function.