Thesis
Classifying agent failures by symptoms alone is not enough. The taxonomy should also describe observability gaps, containment controls, and reproducibility requirements.
Outline
- Planning failures
Goal drift, invalid decomposition, missing constraints, and brittle long-horizon execution.
- Tool-use failures
Unsafe calls, malformed actions, permission confusion, and partial side effects.
- Memory failures
Stale context, poisoned memory, privacy leakage, and incorrect recall.
- Coordination failures
Multi-agent handoff errors, conflicting plans, and hidden dependencies.
- Observability mapping
Required traces, metrics, and replay controls for each failure family.
This framework will be developed as a public artifact that can feed future demos and eval harnesses.
References
References will be added from public AI agent, safety eval, security, reliability, and incident taxonomy sources.