Most of the work in vertical agent tooling ends up in shaping the domain APIs, not the model layer. How are you handling errors the model can't recover from? Surfaced as tool errors for retry?