How a Coding Agent Reshaped One Developer's Approach to A/B Test Analytics

A developer building a price-tag scanning screen for a retail Android app faced an indecisive team vote between three interface designs, prompting a live A/B/C test with real store staff. A coding agent generated all three layout variants and the underlying screen code, but the more valuable contribution came in designing the telemetry architecture. Rather than logging dozens of individual user actions, the agent proposed modeling a 'scan session' entity captured by just two events — a session start and a session finish — each carrying rich contextual parameters. This approach allowed the team to analyze complete scan attempts rather than isolated clicks, a shift the developer says he would not have arrived at independently. The project also surfaced practical pitfalls in exporting Firebase Analytics data to BigQuery, including deduplication traps and the limitations of using averages instead of percentile metrics.