Vercel auto-merges 58% of PRs in its largest monorepo without human review, cutting average merge time from 29 hours to 10.9
Vercel's oldest and largest Next.js monorepo (marketing site, docs, sign-up flow, dashboards, internal tooling) gets over 400 PRs a week, and every one needed human approval. Vercel built an LLM risk classifier (via its AI Gateway) that labels each PR HIGH or LOW risk from the diff, title and description, with hard rules for 100+ changed files and CODEOWNERS paths. Vercel Agent then auto-approves LOW-risk PRs. After a three-phase rollout (silent classification, visible labels, enforcement), 58% of PRs merge without a human reviewer and average merge time fell 62%, from 29 hours to 10.9. 671 low-risk PRs skipped review and none were reverted. PRs per active author rose from 2.6 to 3.8 a week. The workflow is now permanent on that repo.
The numbers (as reported)
| PRs merged without a human reviewer | 58% |
|---|---|
| Average merge time | 29 hours → 10.9 hours (-62%) |
| PR volume | Over 400 PRs a week |
| Low-risk PRs that skipped review | 671, zero reverted |
| PRs per active author | 2.6 → 3.8 a week (+46%) |
| Time to first review, HIGH-risk large-diff PRs | 24.7 hours → 9.0 hours (2.7×) |
| Deployment rollbacks | 2.8 → 1.9 a week |
| Classifier cost | ~$0.054 per assessment (~$51/week) |
Facts
- Company / subject
- Vercel
- Tools
- Unstated
- In-house tools
- Vercel Agent, LLM PR risk classifier (Vercel), Vercel AI Gateway
- Codebase
- brownfield
- Work type
- Company rollout
- Stack
- Next.js, GitHub, Datadog
- Published
- 2026-04-06
- Last checked
- 2026-11-10
What the source claims Vendor-published
Claimed by the vendor that sells the tool. Not checked independently.
- An agent merges 58% of PRs without a human reviewer, and average merge time dropped 62% from 29 hours to 10.9 hours 29 hours → 10.9 hours (-62%)
“Today, an agent reviews and merges 58% of those pull requests without a human reviewer, and average merge time has dropped 62%, from 29 hours to 10.9 hours.”
- 671 low-risk PRs skipped review and none were reverted 671 PRs, 0 reverts
“671 low-risk PRs skipped review. Zero were reverted.”
- Individual human throughput up 46%, from 2.6 to 3.8 PRs per author per week 2.6 → 3.8 PRs per author per week
“Individual human throughput increased 46%. PRs per active author went from 2.6 per week to 3.8 per week.”
- Time to first review on HIGH-risk, large-diff PRs fell from 24.7 to 9.0 hours 24.7 hours → 9.0 hours (2.7×)
“Time-to-first-review on HIGH-risk, large-diff PRs dropped from 24.7 hours to 9.0 hours, a 2.7x improvement.”
- Deployment rollbacks fell from 2.8 to 1.9 a week 2.8 → 1.9 rollbacks a week
“Deployment rollbacks decreased from 2.8 per week to 1.9 per week.”
Caveats
Vendor-published: Vercel sells Vercel Agent and AI Gateway, which run the workflow. Dead original link: the original vercel.com address (https://vercel.com/blog/58-percent-of-prs-in-our-largest-monorepo-merge-without-human-review) now returns 404 and is not in Vercel's sitemap, so this story links to the Wayback Machine copy of 7 April 2026 (checked 10 November 2026); all numbers were verified from that archived copy, which shows no author names. The source covers an LLM classifier plus auto-approval, not a coding agent writing the code: it does not say which agents (if any) authored the PRs, and it does not name the classifier model (secondary write-ups say Gemini). Self-reported, from an experiment window of unstated length, in a repo dominated by marketing, docs and UI changes, so transfer to riskier codebases is unproven. The control group had 2 reverts out of 513 reviewed low-risk PRs. Cost figures are per-assessment rates, not a total.
More stories by tool: Unstated
Related topics: Company rollouts and costBrownfield codebasesVendor case studies