Can the frustration classifier tell why a user is upset, or just that they are?

Just that they are: the classifier scores the tone of what someone wrote, not the reason behind it, so it can’t tell you whether they’re upset about a wrong answer, a slow response, or something else entirely.

It judges surface tone: annoyance, anger, or disappointment in the words themselves, not the cause. A flat bug report with no emotional charge in it reads as neutral even when the underlying issue is serious, and someone venting reads as frustrated even if the agent’s answer was correct.

That’s not a gap to fix later, it’s the tradeoff that keeps the classifier cheap enough to run on every trace. Explaining why someone is upset needs the full trace: what the agent said, what tools it called, where the conversation actually went wrong. The classifier’s job is narrower, to notice the tone fast enough that the trace reaches triage instead of disappearing into traffic that looks fine by every other metric.