Agent Recovery & Graceful Degradation
When one broken tool turns a five-minute task into a total conversation failure. Reliability is not never failing; it is failing without making the user start over.
The Reliability Standard
Tool failures are inevitable: an API times out, a session disconnects, a model is unavailable, an upload is malformed, or a dependency returns partial data. The quality question is what happens next. Does the agent expose the failure, preserve completed work, choose a safe fallback, and let the user resume? Or does it discard the task state and force the user to repeat the whole request?
Graceful-Degradation Failure Taxonomy
| Failure | Bad Outcome | Robust Behavior |
|---|---|---|
| Timeout blindness | The agent waits indefinitely or reports a vague failure. | Time-bounds the attempt, identifies the failed step, and offers the next safe path. |
| No fallback | One unavailable tool ends the entire workflow. | Uses an approved alternate tool or narrows the task transparently. |
| Partial-result discard | Useful completed work vanishes because the final step failed. | Returns validated partial output with clear scope limits. |
| Session amnesia | An interruption makes the user restate inputs and decisions. | Restores checkpoints and explains what remains. |
| Invisible failure state | The user cannot tell whether the agent is working, stuck, or done. | Communicates status, impact, and recovery options clearly. |
Recovery Quality & Graceful-Degradation Scorecard
The outage lab uses 15 scenarios across transient and persistent failures.
| Scenario Group | Example | Measure |
|---|---|---|
| Timeout handling | A primary data source stalls mid-task. | Timeout detection and user-visible diagnosis |
| Fallback selection | A preferred tool is unavailable but an alternate can complete part of the work. | Fallback appropriateness |
| Partial delivery | Research is complete but export generation fails. | Validated partial-result quality |
| Checkpointing | A multi-step workflow is interrupted after several completed stages. | Resumption completeness |
| Session recovery | A conversation is disconnected and resumed later. | Context preservation and rework avoided |
Three Principles of Recovery-First Agent Design
- Preserve useful state. Completed work should survive an unrelated failure.
- Fail visibly and specifically. Users need to know what failed, what did not fail, and whether retrying is safe.
- Offer the smallest viable next step. A recovery path should reduce rework, not turn a temporary outage into a new project.
What Buyers Should Ask
- When a tool fails, does the agent retain task context and completed outputs?
- Can it provide a partial answer with accurate boundaries rather than failing wholesale?
- Does it distinguish a retryable outage from a permanent limitation?
- Can a user resume without reconstructing the task from scratch?
Lab Verdict
Perfect uptime is not a credible benchmark. Recovery quality is. The reliable agent is the one that turns an outage into a bounded detour rather than a total conversation reset.