What happens when a workflow breaks?

It stops and comes to you. A failed run on a codified workflow lands in your inbox as a failure review, not as a log line somebody has to go looking for.

From there the failure opens a repair interview: the same authoring surface the workflow was built in, pointed at the run that just failed. The repair agent reads the real artifacts, the failed run, the step's inputs, the code and the examples the workflow already handles, and then recommends one of two things in the open.

A patch fixes a changed field or an edge case nobody had seen. It gets checked against the known examples first, so the fix cannot quietly break what already worked. You confirm it and the workflow carries on.

A demotion sends the workflow back to learning, for when an assumption broke rather than a detail. It keeps everything it learned, handles the new reality agentically with you in the loop, and collects fresh examples of the part that changed. Work arriving meanwhile queues instead of failing. Once you promote the repaired version, the queue drains through it, so a hundred backed up events do not mean a hundred approvals.

There is deliberately no machine verdict grading how serious the failure was. The agent that can open the run, the code and the examples side by side reads a failure better than a classifier that sees a summary of it, and a recommendation you can overrule beats a decision you never see.