Failure taxonomy
Mapping each failure to its mitigation is how you turn agent reliability into an engineering problem.
- Data quality — fix parsing, ACL propagation, contextual chunk prefixes.
- Tool sprawl — cap at ~10 tools per agent; decompose into sub-agents.
- Eval rot — sample production failures into the golden set weekly.
- Prompt drift — version prompts, run regression suite on every change.
- Permission leaks — query-time ACL filters, never trust embeddings alone.
- Cost runaway — token budgets per task, cached retrieval, model routing.
- Hallucinated tool calls — strict JSON schemas, tool result validation.
- Weak HITL — explicit approval gates for irreversible actions.