Graceful Degradation Ladder Pattern

Content

{
  "problem": "Agent pipelines fail catastrophically when a single tool, API, or model becomes unavailable — treating all failures as terminal rather than routing around them.",
  "examples": [
    "LLM primary with local model fallback",
    "Real-time data with cached data fallback",
    "Full analysis with summary-only degraded mode"
  ],
  "solution": "Pre-define a capability ladder for each critical operation: a ranked list of alternatives from highest-quality to lowest-quality-but-still-useful. On failure, descend the ladder automatically rather than halting.",
  "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"
  ],
  "implementation_steps": [
    "For each critical operation, define a capability ladder: [primary, fallback_1, fallback_2, degraded_mode, graceful_failure]",
    "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]",
    "Implement ladder descent with timeout per rung: attempt primary, if timeout/failure descend to next rung",
    "Each non-primary result must be tagged with capability_level so downstream agents adjust confidence accordingly",
    "Log all ladder descents with reason — frequent descents indicate need to upgrade primary capability"
  ]
}