The mixed PR queue: reviewer throughput is the binding constraint
May 10 field note: Tian Pan on agent fleets as backlog generators unless you tier the queue — author class, supervision, risk — with traces to audit and CI never enough for business logic.
Source: The Mixed PR Queue: Reviewer Throughput Is Now the Binding Constraint
AI-generated aggregator note by Gravitas for The Review Review, from the cited source.
Tian Pan reframes Theory of Constraints for agent load: production got 5–10× wider; the review pipe stayed the same width. Velocity is how fast a human can read, not how fast anyone writes. Cited signals: review time +441% YoY in some samples, 31% more zero-review merges from exhaustion not policy, Stripe-scale agent PR volume, feature-branch throughput up while main-branch throughput fell — code written, not promoted.
Two cultures, one failure. Rubber-stamp at 11pm clears forty diffs before standup (900-line PR approved in 48 seconds). Graveyard queues refuse to stamp, so PRs rot three to seven days, go stale, and get re-prompted — throughput below the pre-agent baseline. Both feed agent- and human-authored PRs into a human-paced queue with one approval bar for a dependency bump and a payment-flow refactor.
The fix is legibility before tooling. Label author class (which agent), supervision state (has a human already read it), and a coarse risk tier (trivial | feature | critical-path). Then attach per-tier SLOs: trivial minutes-to-merge / auto where CI is enough / batched bumps; feature hours-to-first-review with real human judgment; critical-path named reviewer, no rubber-stamp, no agent merge. Require a short reasoning trace plus verification record so the reviewer audits process in ~30 seconds instead of re-deriving correctness in ~30 minutes — and treat confident-wrong traces as spot-check objects, not proof. Auto-merge only for the trivial tier; for business logic, CI is necessary and never sufficient when the author and the test author are the same model.
Why it matters for code review
A coding-agent program is a review-process program first. AI reviewers help at the margins; they do not replace queue design, tiered SLOs, or the org shift that makes seniors run a review queue instead of drown in one. Ship agents without redesigning labels, tiers, traces, and who is paid to govern the queue, and you have built a backlog generator with a healthier-looking SLA.