How do you test tool-calling reliability before shipping?

Been wiring up an agent that hits a few internal APIs, and the failure modes only show up under load: timeouts that trigger duplicate writes, and the model quietly picking the wrong tool when two schemas overlap. Curious how others handle this. Do you run offline evals against recorded tool traces, or just lean on retries and idempotency keys? And what are you using for tracing — LangSmith, OpenTelemetry, something homegrown? Right now I log every call with inputs and outputs and diff against a golden set, but it