Graceful Degradation

The Core Principle

Graceful degradation means a system's failure mode is *partial, transparent, and recoverable* rather than *total and opaque*. In multi-agent pipelines, this translates to a specific design contract: when a subagent fails, times out, or returns unusable output, the coordinator does not propagate that failure as a top-level error. Instead, it returns the results it *does* have, annotates what is missing, and explains why.

This principle is fundamentally about user trust. A system that returns 80% of a result with a clear explanation that 20% is unavailable is more useful — and more trustworthy — than a system that silently hangs, returns a generic error, or waits indefinitely for a timeout.

The Four Design Decisions

1. Define the minimum viable result. Before building, specify what a degraded-but-acceptable response looks like for each pipeline. A financial summary pipeline might degrade acceptably if the news-sentiment subagent fails, but not if the account-data subagent fails.

2. Classify dependencies. Some subagent outputs are *blocking* (downstream agents cannot run without them) and some are *enriching* (they add value but are not required). Graceful degradation only applies to enriching dependencies. Blocking dependencies still require either a retry or an escalation pattern.

3. Design explicit partial-result types. Use structured output with an optional field pattern:

{
  "account_summary": { ... },    // present
  "news_sentiment": null,        // absent
  "degradation_notes": [
    "news_sentiment: subagent timeout after 30s — sentiment analysis unavailable"
  ]
}

4. Surface degradation to the user. The user must know what they are not getting. A degraded result presented as complete is a trust violation. The degradation note should say what is missing and, where possible, when it will be available (e.g., retry later, request human review).

How It Differs From Error Handling

Exam Trap: Silent Degradation

A common exam distractor is a coordinator that drops a failed subagent's output and returns the remaining results *without any note*. This is not graceful degradation — it is silent data loss. Graceful degradation requires transparency. If the coordinator's response contains no indication that something was omitted, the design fails the pattern regardless of whether it returns a partial result.

For the related pattern of recognizing when even partial results are insufficient, see Escalation Pattern. For how individual subagent failures should communicate, see error-propagation. The concept guide Graceful Degradation with Transparency covers how to write the degradation notes users actually need.

Question Angles

Scenario identification: A coordinator receives a timeout from one of three subagents. Four response options are presented — the correct one returns partial results with a degradation note; distractors include: returning a full error, silently omitting the missing data, or indefinitely retrying. You must identify the graceful-degradation response.

Design critique: A pipeline design is shown. You are asked which subagent failure would or would not be suitable for graceful degradation. Blocking dependencies (the failure invalidates all other outputs) are not suitable.

Structured output shape: You are given a partial JSON response and asked to identify whether it meets graceful degradation requirements. Watch for the degradation_notes / missing_fields annotation — absence of it disqualifies the response.

Common Confusions

• Graceful degradation vs. retry: Retrying is appropriate when the failure is transient and the result is blocking. Graceful degradation is for non-blocking enrichment failures where returning a partial result is acceptable.
• Graceful degradation vs. fallback: A fallback substitutes an alternative value (e.g., a cached result). Graceful degradation returns *nothing* for the failed component but is transparent about it. Both can coexist.
• "Degraded" means the result is wrong: False. A degraded result is a *complete subset* of the full result, not a corrupted one. The present fields are as accurate as if all subagents had succeeded.