What It Actually Takes to Run an AI Agent in Production

I took a single-purpose customer support agent (read a ticket, search the runbooks, query the logs, write an incident summary) and pushed it through the Gemini Enterprise Agent Platform on Google Cloud to fix all four. Including a bug I planted on purpose to show why a single eval score hides the most dangerous class of retrieval failure. 📦 FULL PROJECT ON GITHUB: agent scaffold, evals, deployment config, registry setup. Clone it and build the same thing yourself: https://github.com/bytemonk-a