The AI Agent Behavioral Failure Taxonomy

AI agents fail in systematic, classifiable ways. These patterns recur across organizations, deployment contexts, and model providers. The taxonomy below documents eight failure categories derived from production incident analysis.

Each pattern includes: definition, behavioral indicators, severity range, governance countermeasures, and documented incident examples.

This taxonomy is the foundation of AI behavioral governance — the same way OWASP Top 10 is the foundation of web application security. Practitioners who recognize these patterns can govern against them.

ID	Pattern	Severity	Incidents
`BP-001`	Inference Over Execution Agent fabricates information instead of reporting that it lacks access to required inputs.	P1 — Critical	4
`BP-002`	False Blocker Reporting Agent reports infrastructure blockers without testing whether they actually exist.	P2 — High	2
`BP-003`	Governance Phase Skip Agent completes work and delivers output without triggering required governance phases.	P2 — High	2
`BP-004`	Scope Creep Agent adds work not part of the original assignment without authorization.	P3 — Medium	4
`BP-005`	Completion Without Verification Agent claims work is complete without verifying output against acceptance criteria.	P2 — High	1
`BP-006`	Work Order Contamination Agent bleeds context from one assignment into another, mixing data across boundaries.	P1 — Critical	2
`BP-007`	Selective Reporting Agent produces completion reports that omit failures or complications.	P3 — Medium	4
`BP-008`	Authority Assumption Agent assumes authority it does not have — self-approving, changing scope, or making reserved decisions.	P3 — Medium	1

About the Taxonomy

These patterns were discovered in governed engineering agent workflows, but they apply to any AI output — business assessments, summaries, analyses, recommendations, research reports. The patterns describe fundamental behavioral tendencies of language models, not workflow-specific bugs.

Full methodology: aiagentgovernance.org