LEAD-01-043-P · Primer — The Failure Taxonomy

The idea in one sentence

Agents fail in a small number of recognizable, recurring ways — and naming each mode is the first step to controlling it.

Why it exists

You cannot manage a risk you cannot name, and 'the AI messed up' is not a category a leader can act on. Just as an airline maintains a taxonomy of failure types so each has an inspection and a checklist, an agent program needs a shared vocabulary for how things go wrong. Without it, every incident feels novel and no control ever gets built — with it, each failure maps to a known mitigation.

How it works

Six modes cover the vast majority of agent failures. (1) Hallucination — the model states something false with confidence, e.g. inventing a refund policy that does not exist. (2) Tool failure — an API times out or returns garbage and the agent proceeds as if it succeeded. (3) Infinite loops — the agent retries or re-plans forever, burning cost and time. (4) Bad planning — it picks a valid-looking but wrong sequence of steps toward the goal. (5) Memory drift — accumulated context contradicts itself and the agent acts on the wrong version. (6) Stale knowledge — it acts on out-of-date facts, like last quarter's pricing. The discipline is to instrument for each: a mismatch between claimed and actual tool results catches tool failure; a step-count ceiling catches loops. In a finance-close agent, hallucination looks like a fabricated journal entry, while stale knowledge looks like using an expired FX rate — same agent, two different controls.

See it

flowchart TD
  A["Agent action"] --> B["Hallucination<br/>confident but false"]
  A --> C["Tool failure<br/>bad or silent result"]
  A --> D["Infinite loop<br/>never terminates"]
  A --> E["Bad planning<br/>wrong step order"]
  A --> F["Memory drift<br/>context conflicts"]
  A --> G["Stale knowledge<br/>out-of-date facts"]

In practice

A claims-processing agent quotes a policy exclusion that was removed six months ago (stale knowledge), then, when the customer disputes it, retries the lookup nine times against a dead endpoint (tool failure + loop) before approving the claim on a fabricated rationale (hallucination). One incident, three named modes, three separate controls.

The one thing to remember

If you can't name the failure mode, you can't build the control — six categories cover almost everything.