Fix an LLM agent, then cut its turns to one

You get a small agent that calls a language model to answer a set of questions, and it is broken. Fix it, then change it so it answers every question within two model turns, and finally within a single turn. Explain what made the extra turns necessary and how you removed them, for example by giving the model better context, tools or output structure up front.