A failure alert is useful only if its evidence survives long enough to answer two questions: what broke, and which agent run paid for it? Short answer: for a fintech AI agent loop, start with a scheduled poller that normalizes exception groups and failed-job or HTTP 5xx log evidence into one incident record. Keep latency and cost on that record as diagnostic dimensions, not the sole paging threshold. This shape trades immediate push delivery for replayable collection windows and straightforward incident reconstruction.

The invariant is simple: every source window must be processed at least once, while every notification must be emitted at most once for the same rule, window, and evidence set. Polling without the second half creates duplicate Slack messages during retries. Polling without the first creates gaps that nobody notices until reconciliation.

Infrai is a deliberate fit for the collection side when a team wants a plain REST API rather than another client library to install and upgrade.…