30 Days as the Only Human in an AI Dev Team: 212 PRs, Broken CI

A month-long experiment building a multi-tenant SaaS product almost entirely with AI coding agents surfaced a stark lesson: when agents can produce work faster than humans can verify it, every human-paced control degrades to zero while dashboards keep reporting green. Over 30 days, agents merged 212 pull requests and changed more than 111,000 lines of code, but 208 of those PRs received no human review, with a median time-to-merge of about 90 seconds.

The root cause wasn't agent dishonesty — agents flagged their own shortcuts in PR bodies and code comments — but a verification vacuum: self-reported 'green' status went unchecked. A broken CI step went undetected for 16 days, during which roughly 25 PRs merged over a red gates check, and the staging environment served Terraform's placeholder image the entire month despite dashboards showing successful deploys.

The findings reframe code review as a scarce resource rather than a blanket practice: one engineer can properly review around 30 high-risk PRs a month, not 212. The proposed fix is risk-tiered sign-off, machine-enforced verification independent of agent self-reporting, and automatically converting disclosed technical debt into tracked work — since disclosure without consequence proved worthless.