Limits of automated retry

Definition

Retrying with validation-error feedback fixes most failures fast — but not all — so an agent must recognize when to fail fast and escalate instead of looping.

Key points

• The pattern: append specific validation errors to the prompt and retry; this resolves most failures in 2-3 attempts.
• Effective for formatting errors — nested objects vs flat arrays, locale-formatted strings; a retry loop (generate → validate → success, feeding the error message back on failure) fixes these.
• Ineffective for missing information — the loop hits Max Retries Exceeded. Example: extracting a full author list when the source says "et al." and points to an unprovided external document.
• Recognize when to fail fast — inability to make meaningful progress is one of the three valid escalation triggers.
• Ties to structured errors: validation/permission categories are not retryable (isRetryable:false); only transient errors are — see Structured tool errors.

Why it matters for the exam

• "The agent keeps retrying a missing-data failure" scenarios test recognizing that retry won't help and escalation/fail-fast is correct.

Common gotchas

• Missing information is the classic case where retry is least effective — don't keep looping.
• 2-3 attempts is the rule of thumb; endless retries are an anti-pattern.

See also

• Escalation triggers
• Structured Tool Errors
• Graceful Tool Failure
• Minimal footprint principle

Sources

• architects-playbook
• netlight-study-guide