Ask HN: How do you debug multi-step AI workflows when the output is wrong?
I’ve been building multi-step AI workflows with multiple agents (planning, reasoning, tool use, etc.), and I sometimes run into cases where the final output is incorrect even though nothing technically fails. There are no runtime errors - just wrong results.The main challenge is figuring out where things went wrong. The issue could be in an early reasoning step, how context is passed between...