I've been running a few multi-step agent workflows in production and the biggest reliability headache isn't the model, it's retry logic. A tool call times out, the agent retries, and now we've got duplicate side effects — two calendar events, two charged invoices. Idempotency keys help, but every integration seems to handle them differently.
Curious how people are structuring this. Are you wrapping tools in your own retry layer, leaning on the framework's built-in handling, or just making every action idempotent at the API level? Also interested in whether anyone's found decent observability tooling for tracing where a run actually failed versus where it just looked like it failed.
Curious how people are structuring this. Are you wrapping tools in your own retry layer, leaning on the framework's built-in handling, or just making every action idempotent at the API level? Also interested in whether anyone's found decent observability tooling for tracing where a run actually failed versus where it just looked like it failed.