What's the hardest part of building reliable AI agents?

[JSONFIRST]

We built JSONFIRST to solve one specific problem: ambiguous user input breaking agent pipelines.

But there are many other hard problems in building production AI agents:

• Hallucinations on edge cases
• Multi-step reasoning failures
• Routing the right task to the right agent
• Handling low-confidence outputs safely
• Debugging failures in production

What's the hardest part for you? What have you tried that didn't work?

Looking to understand real pain points in the community — all feedback helps.