Meta's AI coding tools have outpaced the humans reviewing their output, so the company automated the review layer too.
At Meta, AI-assisted coding has pushed code volumes to extremes. Lines of code per landed diff grew 105.9% year over year; per-developer diff volume rose 51%; and agentic AI drove more than 80% of that growth. The surge created a straightforward bottleneck: fewer diffs were receiving timely human review. Meta's answer was RADAR (Risk Aware Diff Auto Review), a multi-stage pipeline that scores each diff for risk using static heuristics, a machine-learned model, and LLM-based review before deciding whether to land the change automatically. Of 535,000-plus diffs RADAR has evaluated, more than 331,000 were landed without a human approver.
The safety numbers Meta reports are harder to dismiss than the speed ones. RADAR-reviewed diffs revert at one-third the rate of non-RADAR diffs, and production incidents are one-fiftieth as frequent; the system also cuts median time-to-close by more than 330%. But RADAR is explicitly designed to cherry-pick low-risk changes, so those comparisons aren't apples-to-apples: non-RADAR diffs include the riskier code the system declined to touch.
Every organization running AI coding tools at scale will hit this same arithmetic, and the real question is whether RADAR-style automation stays inside hyperscalers with the telemetry depth to tune it, or becomes a pattern smaller teams can credibly adopt without 535,000 training diffs.