Graceful Degradation Ladder Pattern

Problem

Solution

Implementation steps

1. For each critical operation, define a capability ladder: [primary, fallback_1, fallback_2, degraded_mode, graceful_failure]
2. Example ladder for web search: [Exa neural search] → [Tavily keyword search] → [Google Custom Search API] → [LLM knowledge only, flag as unverified] → [Return "search unavailable" with partial answer]
3. Implement ladder descent with timeout per rung: attempt primary, if timeout/failure descend to next rung
4. Each non-primary result must be tagged with capability_level so downstream agents adjust confidence accordingly
5. Log all ladder descents with reason — frequent descents indicate need to upgrade primary capability

Examples

• –LLM primary with local model fallback
• –Real-time data with cached data fallback
• –Full analysis with summary-only degraded mode

Anti-patterns

• –Infinitely retrying the same failed operation instead of descending the ladder
• –Hiding capability degradation from downstream agents — they need to adjust confidence
• –Building ladders so wide that degraded output is misleadingly similar to primary quality