Fallbacks and model routing

Have a fallback path for when the primary model or route fails, and send each task to the model best suited to it.

Always have a plan B

Providers have outages, models get rate limited, and some requests just fail. A fallback, whether a second provider, a smaller model, or a simpler deterministic path, keeps the agent working when the primary route is unavailable. Reliability often comes down to whether there is a plan B at all.

Route to the right model

Not every step needs the largest model. Routing simple steps to a fast, cheap model and reserving the strongest model for the hard ones improves both cost and reliability, because each step runs on something well matched to it.

Fail down, not out

When the ideal path is unavailable, degrade to a lesser but working one rather than failing entirely. A slightly worse answer delivered reliably usually beats the best answer delivered sometimes.