Skip to content
AX Framework
Pillar 06 of 06

Recovery & Resilience

Can an agent handle errors, ambiguity, and interruptions?

AX proposal

Failures are specific, recoverable, and safe to retry, so an interrupted task can resume rather than restart.

Agents encounter more failure than humans do. They run longer sessions, chain more steps, and hit rate limits, expired tokens, and changed inventory. The question is not whether things go wrong but whether the product helps the agent recover.

Recovery & Resilience asks whether errors explain what went wrong and what to do next, whether retries are safe, whether partial work can be rolled back, and whether a task can resume after an interruption without losing context.

A resilient product treats an error as a turn in the conversation rather than the end of it.

What good looks like

Signals

  • Errors carry stable codes, human-readable messages, and suggested next actions.
  • Retryable errors are marked as such, with timing guidance.
  • Completed steps of a workflow can be compensated or rolled back.
  • Sessions and workflows can be resumed from a token or identifier.
  • Ambiguous requests return clarification options instead of guessing.

Anti-patterns

  • Generic "Something went wrong" errors.
  • Retries that duplicate side effects.
  • Workflows that must restart from the beginning after any failure.
  • Silent fallbacks that change the meaning of a request.

Evaluation checks

Use these questions when reviewing a task. They feed the AX evaluation.

  1. R1Do errors include a stable code and a suggested recovery action?
  2. R2Are retryable and non-retryable errors distinguished?
  3. R3Can a partially completed workflow be resumed or rolled back?
  4. R4Does the product ask for clarification when a request is ambiguous?