Errors, retries & alerts

A workflow that ran perfectly all week will, eventually, meet reality: the sheets service has a bad minute, an email arrives in a format nobody imagined, the AI step times out mid-morning rush. Unreliable automations and reliable ones meet the same failures — the reliable ones were built expecting them. Three mechanisms cover nearly everything:

• Retries, for transient failures. Most API errors are weather — down for seconds, fine on the second try. Most platforms offer retries — some automatic, some enabled per step or via an error-handler route; find yours and turn it on for every step that calls an external service. The one caution worth learning early: retry reads freely, but retry a send only if you're sure the first attempt didn't land — double-charging and double-emailing are retry bugs. When unsure, fail to the problem folder instead.
• The problem folder, for persistent failures. When a run fails past retries or data won't parse, the run's details land in one designated place — a 'Failed runs' sheet tab, a triage channel — with what came in, which step failed, and the error. This is the unsure lane's sibling (engineers call it a dead-letter queue): nothing is silently lost, everything weird has an address, and Monday's fix-it session has a worklist instead of a mystery.
• Alerts, for humans — tiered or ignored. A failed run posts to the problem folder quietly. Repeated failures (3+ in an hour: something systemic) ping the owner directly. The tiering is the point: alert on everything and alerts become wallpaper within a week; alert on patterns and a ping still means something. Silence-as-health only works if failures reliably make noise — that's what you're building.

The pre-mortem: five minutes that finds most of it

Before any workflow goes live, walk it step by step asking two questions: 'what if this step fails?' and 'what if this step's input is garbage?' Write one line per step. Riley's v2 pre-mortem found: AI step down → retries, then problem folder (don't guess a category); malformed AI output → already validated, unsure lane; sheet full/renamed → problem folder + owner ping; chat down → ticket still logs, ping retries (ordering matters: log first, notify second, so notification failures never lose data). Ten minutes, four decisions, zero incidents later — versus discovering each answer live, one bad morning at a time.