COMPEL Glossary / agentic-failure-taxonomy
Agentic Failure Taxonomy
An agentic failure taxonomy is a structured classification system that categorizes the types of failures that can occur in agentic AI systems, providing a shared vocabulary for identifying, discussing, and governing AI agent risks.
What this means in practice
Categories typically include goal misalignment (agent pursues wrong objectives), tool misuse (agent uses authorized tools inappropriately), cascading errors (agent propagates upstream failures), unauthorized escalation (agent exceeds delegated authority), resource overconsumption (agent generates excessive costs), and emergent misbehavior (agents develop unexpected interaction patterns). For organizations deploying agentic AI, a taxonomy enables systematic risk assessment, targeted controls, and effective incident classification. In COMPEL, the agentic failure taxonomy is introduced in Module 3.4, Article 12 on agentic AI risk taxonomy and enterprise risk framework extension.
Why it matters
Without a structured classification of how agentic AI systems can fail, organizations cannot systematically assess risks, design targeted controls, or effectively classify incidents when they occur. A shared vocabulary for agent failure modes enables teams to communicate precisely about risks and ensures that governance controls address the full spectrum of potential failures rather than focusing only on the most obvious ones.
How COMPEL uses it
The agentic failure taxonomy is introduced during the Model stage when designing the enterprise risk framework extension for agentic AI within the Governance pillar. Categories including goal misalignment, tool misuse, cascading errors, and unauthorized escalation are mapped to specific controls. During Evaluate, incidents are classified using the taxonomy to enable pattern recognition, and the Learn stage uses classified incident data to refine governance controls.
Related Terms
Other glossary terms mentioned in this entry's definition and context.