Validation catches structural problems before deployment. It can't catch a workflow that's structurally perfect but fails at runtime — a webhook trigger that never fires the way you expected, an API that rejects a correctly-shaped request for its own reasons. For that, the only real signal is actually running it.
Our self-healing loop deploys a workflow, triggers a real execution, reads back what actually happened, and if it failed, feeds the actual error back into a targeted fix — not a vague "fix this workflow" re-prompt, but a specific instruction naming the exact validation errors found.
It runs in two modes: suggest, which proposes a fix for a human to approve before anything changes, and auto, which applies and re-tests up to a capped number of iterations. Suggest is the default for anything that could trigger a real side effect, like an actual Slack message or email — the autonomy level should match the blast radius of getting it wrong.
