Agentic Failure Modes¶
What It Is¶
The recurring, categorizable ways a marketing agent's output goes wrong — documented with detection signal, severity, and mitigation so they can be designed against. The reason they matter more than ordinary mistakes: automation makes errors scale. One unscrubbed opt-out becomes 10,000 violations; one drifted prompt degrades every subsequent piece.
The Canonical Modes¶
- Hallucination — confident but false claims, fake statistics, invented customers/features. Detection: LLM-judge claim verification (flag confidence < 0.85), statistic-must-match-KB rules, CRM validation of named customers. Severity peaks at Critical for fabricated testimonials (legal liability). Mitigation: internal-KB grounding, citation requirements, unpublish-and-correct workflow.
- Brand-voice drift — output gradually diverges from brand voice over cycles. Detection: LLM-judge score vs. a 30-day gold-standard baseline (flag > 10% decline), terminology/tone tracking. Mitigation: always load the brand-voice doc + best-performing examples as context.
- Over-automation — automating interactions that need human judgment. Detection: engagement/quality decline, customer complaints about "robotic" touch. Mitigation: pull the human back into high-judgment moments.
- Data-quality decay (GIGO) & ICP drift — bad inputs silently corrupt outputs and targeting over time.
How It Applies to Marketing Factory¶
This is the factory's safety catalogue: every agent workflow should map its plausible failure modes to explicit detection checks and a human-review-gate, with the most severe (fabricated testimonials, compliance breaches) treated as red lines that never ship unreviewed. It overlaps the public-input threat of prompt-injection and the specific content risk of brand-voice-drift, and it operationalizes the agent-ownership-boundary: the failure modes mark exactly where autonomy must stop. The unifying control is grounding + detection + gating, because at scale you cannot rely on catching errors after publish.
Related Concepts¶
- human-review-gate — the primary control that catches these before they ship
- prompt-injection — the adversarial failure mode from public inputs
- brand-voice-drift — a specific, high-frequency content failure mode
- agent-ownership-boundary — failure modes define where autonomy must stop
Referenced from: agent-workflow-templates