A failure taxonomy is a classification system for what goes wrong. In multi-agent AI, categories might include incorrect tool use, missing evidence, coordination failure, duplicated work, premature completion or a failure to notice that an earlier step was wrong.
Clear categories make incident reviews more useful. Instead of counting every bad outcome as one generic error, a team can identify the patterns that produce the most failures and map them to changes in prompts, workflow structure or permissions.
A taxonomy needs precise definitions. Two reviewers may disagree about whether a bad result came from reasoning, retrieval or orchestration, especially when several things went wrong together. A single failure may fit more than one category, so coding rules should explain whether multiple labels are allowed.
Publish the category definitions, keep examples for each and check reviewer agreement. Use the taxonomy alongside end-to-end success rates, traces and severity, because a simple count of categories does not tell you how costly each failure is.