Banking customer onboarding used to mean spreadsheets, Slack threads, and copy-pasting Kubernetes and APIM configs between teams and Github repos. My director sketched out an idea to streamline it. I picked it up and built an AI agent to orchestrate the delivery. Using another AI agent.
One agent writing another that touches real banking infrastructure. What could possibly go wrong?
Quite a lot. I'll cover the safety patterns that keep a recursive trust chain under control, the multi-persona approach that replaced the team I don't have, and the production scars that taught me why idempotency matters more than intelligence.
No frameworks pitch. No 30-minute demo. Real code, real banks, real consequences.
