Definition
Graceful recovery (also called “repair” in conversation design literature) is the principle that AI agents should be designed to handle failures gracefully rather than trying to prevent all failures. The question is not how to avoid errors, but how to recover from them when they inevitably occur. People are surprisingly forgiving of AI mistakes if recovery is handled well.
Key Points
- From “Conversations with Things”: human conversations go off the rails constantly, but we fix them together almost unconsciously
- The example: an agent that can’t find “the usual” order should ask naturally (“Could you remind me what that is?”) rather than crash (“I don’t understand ‘the usual’”)
- Recovery is not about canned “sorry” responses — it must be built into every interaction
- Four elements of good recovery: Acknowledgment, Empathy, Humor (when appropriate), Action
- Recovery is more important than perfection — users forgive mistakes if handled gracefully
- This principle applies at every iteration of the agent loop
Related Concepts
- agent-loop — Recovery happens at each loop iteration when an action fails
- conversation-design — Graceful recovery is a core principle of conversation design
- turn-taking — When turn-taking fails, graceful recovery handles the resulting confusion
- context-engineering — Recovery strategies depend on the context available to the agent
Implications
This matters because it shifts the engineering priority from “make it never fail” to “make it recover well.” This is more achievable and more impactful — a 95% accurate agent with great recovery feels better than a 99% accurate agent that crashes on the 1%. It also explains why some technically inferior agents outperform superior ones in user satisfaction.
Open Questions
- How do you measure recovery quality? Is there a “recovery success rate” metric?
- Can recovery patterns be automated, or do they require per-domain customization?
- How does graceful recovery interact with safety guardrails — when should an agent refuse vs. recover?