Retry Queues for Agent Email

Failures that need a queue

Not every failure should retry. A temporary provider error can retry. A policy block, invalid recipient, or missing approval should stop and route for review. Classifying failure types keeps agents from repeating unsafe work.

Backoff schedules

Backoff should slow retries without losing the user intent behind the message. Agent sends often need a maximum age because context can expire. A reply drafted from yesterday data may be wrong today.

Job fields

Each queued send needs enough context to retry safely and enough metadata to explain what happened later.

• ok idempotency key
• ok recipient hash or internal recipient ID
• ok template or draft ID
• ok agent task ID
• ok attempt count and last error
• ok next run timestamp
• ok approval state and audit log pointer

Dead letters and replay

Dead-letter queues should be inspectable by humans. Replay should create a new attempt record while preserving the original failure trail.

• ok Store the final failure reason and provider response.
• ok Mark whether replay is allowed.
• ok Require fresh approval for old outbound replies.
• ok Keep the original idempotency key for duplicate-send checks.