Is this normal? My first production AI agent has humbled me

Is this normal, or am I just bad at this? 😅

Building my first production-level AI agent — a WhatsApp booking bot for a dermat clinic (n8n + AI Agent + Postgres/Redis + EkaCare integration). Booking, reschedule, cancel, the whole flow.

The last couple weeks have been a real lesson in how many ways things can quietly break:

- Agent confidently told a patient "you're booked!" — except the booking had actually failed, it just never found out

- Session memory that worked fine... until a stray character in the key made it forget everyone was ever there

- A field I *knew* was an array turned out to be a stringified array three nodes downstream, and everything silently broke

- A workflow crash mid-booking left the system in this weird limbo state where the agent kept insisting it was "still checking" — forever

None of these showed up until real (test) conversations hit them. Everything looked fine on paper.

So — genuine question for anyone who's shipped agentic workflows to production: is this just what building these systems looks like? And if you've been through this — what's actually helped you catch this stuff earlier? How do you test for the weird edge cases (crashes mid-flow, race conditions, data type mismatches) before a real user finds them for you?

Would love to hear how others approach testing here — feels like there's no "just write unit tests" playbook for this yet, or maybe I just haven't found it. 🙂

2 comments

Is this normal? My first production AI agent has humbled me