LangChain agent reliability — what breaks in production?

[JSONFIRST]

LangChain works well in demos but production agents often fail on edge cases. Common issues:

• Tool selection errors when user intent is ambiguous
• Prompt sensitivity — small wording changes cause different routes
• Hard to test systematically

We added JSONFIRST as a pre-processing step to normalize input before chains run. Curious what reliability issues others have hit and how you solved them.