agent.ts — 429 retry path ignores Retry-After, coordinated amplification under shared provider contention

[SirBrenton]

Finding

The core retry path in agent.ts handles 429 responses with local exponential backoff but does not consult the Retry-After header.

// Local exponential backoff — no Retry-After consulted
Math.min(1000 * 2 ** attempt, 10000)

Affected lines: 2039, 3267, 5884

isRetryableError() defaults to true, so all 429s enter the same retry path regardless of whether the signal is transient throttling (WAIT) or quota exhaustion (STOP).

The consequence

Under shared provider contention, independently-running agents built on VoltAgent can converge their retry windows instead of dispersing them.

Because this sits in the framework retry layer, every downstream agent inherits the behavior automatically. This stays invisible in light usage and surfaces during provider stress or concurrency spikes.

Expected behavior

• Read Retry-After when present
• Use header value as minimum retry delay
• Distinguish quota exhaustion (STOP) from transient
  throttling (WAIT) — do not retry quota exhaustion
• Fall back to exponential backoff only when header absent

Suggested fix shape

if (response.status === 429) {
  const retryAfter = response.headers?.get('retry-after');
  if (retryAfter) {
    const parsed = parseInt(retryAfter, 10);
    if (!isNaN(parsed) && parsed > 0) {
      await new Promise(resolve => setTimeout(resolve, parsed * 1000));
      continue;
    }
  }
  // Fall back to exponential backoff
  await new Promise(resolve =>
    setTimeout(resolve, Math.min(1000 * 2 ** attempt, 10000))
  );
}

Related pattern

retry-after-ignored-under-concurrency

Corpus reference: https://github.com/SirBrenton/pitstop-truth