Agent failure is not merely stale training data. The classes are: epistemic (stale knowledge → retrieval/version pinning); syntactic (malformed calls → JSON Schema/structured output); semantic (schema-valid but wrong → semantic grounding/evals); operational (timeouts, duplicates → retries/idempotency); security (prompt injection, poisoned descriptions → trusted registries/sandboxing); authority (excess scope, confused deputy → least privilege/scoped tokens); and governance (no audit/consent → provenance, policy logging).
Terms
Glossary terms used here: Confused deputy, Consent, Idempotency, Least privilege, Prompt injection, Provenance, Semantic grounding, Structured output.