# Failure Taxonomy

> A named map of how agents break: exfiltration, corruption, cost blowup, loop, drift, silent-wrong. You cannot monitor what you cannot name.

Category: evals

A failure taxonomy is the catalog of the distinct ways your agents fail, each one named and made concrete. It is the vocabulary that the rest of your reliability work is built on.

## Why it matters
Teams talk about agent failure as one fuzzy blob. It is not. A support agent leaking customer data is a different failure than a coding agent corrupting a module, which is different again from an ops agent turning a minor incident into an outage. If you cannot name them, you cannot detect them, eval them, or fix them.

## In practice
You enumerate the failure modes that actually matter for your agents, with examples and detection signals for each. The taxonomy drives your evals, your alerts, and your red-team attacks. New incidents either fit a known mode or expand the map.

## Where it shows up in my work
The Field Manual is organized around this: an incident taxonomy, the failure modes of dark agents, and the practices that catch them before production does.

---
Source: https://ryanhunter.io/lexicon/failure-taxonomy
Part of the Agent Infrastructure Lexicon by Ryan Hunter. https://ryanhunter.io/lexicon