# What is Fallback Behavior?

Also called degraded mode, safe default.

Fallback behavior is what an agent does when its intended path is unavailable, such as a failed tool, an unreachable model, an exhausted budget, or a result below a confidence threshold. A defined fallback makes failure predictable, whereas an undefined one leaves the agent to improvise, which commonly produces a fabricated answer.

Fallbacks are ordered from most to least preferred. A typical chain tries an alternative tool or data source, then a cached or stale result clearly marked as such, then a partial answer stating what is missing, then a clean failure that hands the task to a person. Each level should be a deliberate choice, not whatever the agent happens to do.

Stating what is unavailable beats substituting a plausible value. An agent that cannot reach a pricing service and returns an old figure without saying so has produced a silently wrong answer, which is more damaging than an explicit statement that the current figure could not be retrieved. Degraded results must be labeled as degraded wherever they are consumed.

The direction of failure should follow what the action does. Read paths can often fail open, returning partial data, because an incomplete report is recoverable. Write paths, payments, external messages, and deletions should fail closed, doing nothing, because a partially executed side effect is often worse than none at all and much harder to unwind.

Fallbacks only work if they are exercised. A path that has never run in testing frequently fails when the primary path finally does, exactly when the system is already degraded. Periodically forcing the primary path to fail, in a controlled setting, is what confirms the fallback works and is what most systems skip.

## Key points

- An ordered chain from alternative source to clean failure
- Label degraded and stale results explicitly
- Reads may fail open; writes and sends fail closed
- Undefined fallbacks invite fabricated answers
- Untested fallback paths tend to fail when first needed

## In practice

A reporting agent's primary metrics API times out. It falls back to yesterday's cached snapshot and labels every figure in the output as of yesterday, naming the outage. Its secondary task, emailing the report to a client list, fails closed instead: no message is sent, and the run escalates, because a client receiving stale numbers without context is worse than no report.

## Related terms

- [Tool Error Recovery](/en/glossary/tool-error-recovery)
- [Guard Condition](/en/glossary/guard-condition)
- [Escalation Policy](/en/glossary/escalation-policy)
- [Confidence Threshold](/en/glossary/confidence-threshold)
- [Agent Loop](/en/glossary/agent-loop)

[Back to the AI Glossary](/en/glossary)