why are ai agents still so hard to debug in production?

feels like the industry figured out how to build ai agents faster than how to understand them.

everyone demos agents.
very few teams can confidently answer:

• why an agent failed

• what changed between runs

• whether quality is improving or regressing

• or if the agent is actually reliable over time

curious how people here are handling this today.

what’s currently the most painful part of running ai agents in production? debugging? evals? monitoring? something else?

love to hear from the PH community.