How do you make your agents more reliable?

Many of the agents being built today rely heavily on tool use, calling APIs, functions, or external services to perform tasks. However, these agents often struggle with reliability: they can hallucinate, select the wrong tools, or pass incorrect arguments. While prompt tuning can help to some extent, it usually leads to oversized and hard-to-maintain models. How are you approaching this challenge and ensuring that your agents use tools accurately and efficiently?