A tool that wakes you up when things break should not itself depend on one thing that can break. If Ketl routed every diagnosis to a single model from a single lab, then that lab's bad afternoon would become your bad afternoon, on the exact day you needed the copilot most.
Not every step needs a frontier model
A diagnosis is not one call. It is several steps, and they do not all need the same horsepower. Deciding severity and grouping duplicate alerts is cheap work that a small, fast model does well. Reasoning about cause across logs, traces, deploys and past incidents is the heavy step, and it earns a frontier model. Routing each step to the right model is how you get quality where it matters and a sane bill everywhere else.
- Triage and deduplication route to a fast, inexpensive model.
- Cross-signal root-cause reasoning routes to a frontier model.
- If a provider is degraded, the router falls to the next candidate whose evidence bar it clears.
Open models, on your own ground
Model-agnostic also means you are not forced onto a hosted API. On the Scale plan, Ketl can route to open models you host yourself, so telemetry never leaves your perimeter. The routing table is the same. Only the endpoints change.
The router is a reliability feature. Labs move fast, and the pager should not notice.
Frontier models will keep leapfrogging each other, and open models will keep closing the gap. A registry and a router mean Ketl can take the best of each as it arrives, without a rewrite and without asking you to care which lab answered.