Appendix C: Failure Quick Map

Symptom	First layer to inspect	What to check
Agent ignores system rule	Context	Is the rule present, early, and not truncated?
Agent invents a tool	Structured output / tools	Is tool allowlist validation enforced?
Tool arguments are wrong	Tool contract / validation	Are schemas and argument policies clear?
Agent loops	Control loop	Are iteration caps and repetition detection active?
Context overflow	Memory / context	Is summarization or truncation triggered early?
Unsafe action happens	Guardrails / permissions	Did a deterministic boundary fail?
Tool result is useless	Tool design	Does the observation give enough actionable state?
Final answer is unsupported	Final-answer verification	Is completion evidence required?
Cannot debug run	Observability	Are run IDs, spans, raw outputs, and stop reasons recorded?
Cost is too high	Loop / context / tools	Are budgets, context size, and tool calls controlled?