An incident tool has to work when everything else is on fire. That is a strange time to depend on exactly one model provider being fast, available and unchanged. So Ember was built model-agnostic from the first commit: a thin routing layer sits between the product and any provider.
What the routing layer does
- Picks a model per task. Ranking causes and drafting a rollback do not need the same model, and the router chooses accordingly.
- Fails over. If a provider is slow or down during an incident, the request routes to another rather than stalling the on-call.
- Respects your boundary. On Scale, teams register their own endpoints, so the reasoning runs on models they control.
The interface is deliberately small. Every model call in Ember goes through one function that takes a task and returns a grounded result. Swapping or adding a provider is a change in one place, not a rewrite.
The best model for an incident is the one that answers, grounded, right now.
Being model-agnostic is not a hedge against picking wrong. It is an acknowledgement that the frontier moves, incidents do not wait, and the team running the pager should not be locked to one vendor's worst day.