Agent frameworks keep failing at the boring parts

Spent the week migrating a workflow between agent frameworks, and the model was never the problem. The failures were retries that re-fired a payment tool, a schema change that silently broke argument parsing, and no visibility into which step actually stalled. Curious what others do here. Do you wrap every tool call in idempotency keys, or is that overkill for read-only tools? And how are you catching schema drift before production: contract tests, pinned versions, or just logs? Also open to observability setups that show per-step latency without drowning