Route new unresolved error groups through a scheduled worker, and page only when those failures threaten the customer support agent's service objective. TL;DR: capture backend exceptions, poll groups, deduplicate notifications with durable state, and let your worker deliver Slack, email, or webhook messages. Keep latency and cost accounting alongside that process: an exception count alone cannot tell you whether the agent completed a useful support response.
The deciding constraint is signal quality. One support request can generate several model calls, tool attempts, and retries, so treating every failed attempt as a fresh incident turns the alert channel into a transcript of the retry loop. A useful alert identifies the unresolved failure, the affected workflow, and the evidence that someone should interrupt their work.
How should an error alerting worker poll unresolved API groups?
Start with the customer-visible outcome: a completed support response within the team's latency objective.…