When an agent does the wrong thing, the tempting fix is a rule: a keyword list, a regex, one more sentence in the prompt for this one case.
I don’t fix agents that way anymore. How an agent behaves comes from four things: its prompt, the context it gets, the tools it can use, and the harness around all of that.
So when an answer is bad, I look at those four. What did the agent see? Which tool was missing? A fix there carries over to cases I haven’t met yet. A patch covers the one case it was written for.