Review time on Salesforce's largest pull requests plateaued and then declined

Reviewer minutes attach to submissions. A dashboard that divides total review time by pull requests opened treats a skimmed-and-closed agent branch as one more data point pulling the average down. The minutes spent on it are gone either way ^[10]. The denominator that tells you what shipping costs is merged changes. The dev.to post headlined "Every agent PR costs review time even if it never merges" argues the two numbers separate exactly when an agent inflates the submission count. On its account, triage is both the fastest-growing part of the work and invisible to any per-PR average ^[13][16].

So a rising submission count pushes the per-pull-request average down on its own, whether or not those pull requests are headed for a merge.

Take the reported 30 percent volume rise and hold reviewer hours flat. Review minutes per line of change fall to about 77 percent of their earlier level, a decline near 23 percent ^[15]. Salesforce's own latency signal says the hours did not keep up: review latency climbed quarter over quarter ^[4].

The other dimension is files. A pull request of 1,000 lines across 20 files averages 50 changed lines a file ^[14], and each file is a separate place to reload the surrounding code. Salesforce reports senior reviewers moving between multiple large AI-assisted changesets daily, and says that raises cognitive overhead even when the reviewer only triages the submission ^[8].

The dev.to post says Salesforce attributes escaped vulnerabilities and architectural regressions to a workflow that no longer supports effective reasoning, not to negligence ^[7]. Separating that from disengagement takes one measurement: the merge rate of large pull requests, bucketed by size, over the same quarters as the review-time curve. If merge rates held while review time flattened, reviewers ran out of capacity. If merge rates fell as diffs grew, reviewers were declining to spend on submissions they had learned to recognize ^[11]. Overload needs more capacity or better triage; disengagement needs tighter gating upstream so agents stop producing work the team throws away ^[12].

Salesforce's figures are internal signals from Salesforce's repositories ^[2]. For the plateau to mean the same thing in another shop, that team would need a comparable share of agent-authored diffs, the same people doing triage and review, and a policy that routes every agent branch through the human queue. The exploration result travels more easily. It comes from a 33-week panel of GitHub developers spanning the introduction of the Copilot Coding Agent, and it finds that AI delegation significantly increases engagement in exploration tasks ^[9].

The dev.to account of Salesforce's post does not include that merge rate or the diff size at which review time turns, so as published the curve supports both readings ^[17].