Every agent PR costs review time — even the ones that never merge

Tess Ainsley (Sep 18) reframes the agentic review bottleneck as an accounting problem, not just a quality one. Salesforce’s Scaling Code Reviews (Appajodu & Boyapati, Jan 29) saw ~30% more code volume, PRs regularly past 20 files / 1,000 lines, rising review latency — and then review time on the largest PRs plateaued and declined. That is usually read as shallower review. Ainsley adds a second reading: if a growing share of large agent PRs never merge, reviewers learn to disengage from unmergeable submissions. Overload and rational triage need different fixes — more capacity vs tighter upstream gating so agents stop opening work the team will throw away.

The metric that hides the cost is review time per submission. Every PR that enters the queue spends reviewer attention whether it merges, gets abandoned, superseded, or rejected. Agents inflate exploration (Tang, Zhao & Karahanna on Copilot Coding Agent delegation), so triage minutes attach to empty denominators. Luo & Guo’s AMCIS 2026 survey of AI-era productivity metrics finds teams still trust volume proxies (LOC, commits, PR count) while delivery stays constrained by review and QA. Ainsley’s fix is concrete: total reviewer time ÷ changes that reached production in the period. Climbing cost-per-merge means the review system is absorbing agent-flow waste; stable cost means the plateau is mostly noise from dead submissions; falling cost is either healthy skip of exploratory junk or dangerous disengagement.

Why it matters for code review

Always-on factories that keep modest verified momentum overnight still lose if daytime dashboards celebrate submission volume. Continuous verification compounds only when the tax you pay is priced against shipped change — otherwise FOMO about idle agents pushes more unverified PRs into a queue that burns seniors whether or not anything lands. Gate upstream; measure the denominator.