Fallback
A predefined alternative action an AI agent takes when its primary approach fails—such as escalating to a human when confidence is low, switching to a simpler model when latency spikes, or returning a canned response when the knowledge base has no match. Well-designed fallbacks prevent agents from failing silently or producing low-quality outputs, and they are the recovery layer of reliability engineering: a fallback is the *specific* alternative action, while graceful degradation is the design principle that there always be a safe one to drop to. Because agents are non-deterministic and retry often, any fallback that re-attempts an action must target an idempotent operation, or the retry itself becomes a new failure.
Written by Max Zeshut
Founder at Agentmelt · Last updated Sep 9, 2026
Example
A support agent can't find a policy answer with high confidence. Rather than guessing, its fallback escalates to a human with the conversation and the low-confidence flag attached — a clean handoff instead of a hallucinated answer.
Frequently asked questions
- What makes a good agent fallback?
- It should be safe (never more risky than the primary action), visible (the degradation is logged, not disguised as success), and matched to the failure — retry with backoff for transient errors on idempotent actions, a simpler model for latency or overload, and a human handoff for anything high-stakes or low-confidence. The goal isn't zero failures; it's zero surprising, silent, unrecoverable ones.