- Meta rolled out RADAR, an automated review pipeline that scores code changes for risk before deciding whether a machine or a human should approve them.
- The system first flags the author and source, runs static heuristics, then applies a learned Diff Risk Score. Low‑risk diffs get an LLM‑generated review and deterministic checks; higher‑risk ones fall back to humans. In five months RADAR examined 535 K diffs and landed 331 K. Raising the risk‑score cutoff from the 25th to the 50th percentile lifted the automatic approve rate to 60.31%.
- Automated reviews reverted at one‑third the rate of manual ones, and production incidents were 1/50 as common. Median time to close a diff dropped by more than 330%, and wall‑time spent waiting for review fell 35%. The gains arrive as Meta’s AI‑generated code output grew 105.9% year‑over‑year, outpacing reviewer capacity.
- The takeaway: layered, risk‑aware automation can keep pace with AI‑driven code growth without sacrificing safety. It shows a viable path for other large developers facing similar bottlenecks, though the benefits hinge on accurate risk scoring and the willingness to trust machines with low‑risk changes.