Your PR Queue Is Out of Control. Vercel Eve + OpenAI Codex Might Fix That.

Vercel's Eve framework — combined with OpenAI Codex — now makes it possible to build a working AI code-review bot in roughly 30 minutes, and the people doing this at scale are already seeing real results. That's the substance of a recent hands-on episode from Claire Vo on Lenny's Newsletter, and it's worth paying attention to even if you've never merged a pull request in your life.

The problem nobody warned you about

Here's the awkward side effect of vibe-coding your way to a working app: every time you (or an AI agent like Devin) generates a chunk of code and pushes it to GitHub, it creates something called a pull request — a proposed change that someone has to review before it goes live. When AI can generate code at 10× the old human pace, the review queue grows faster than any team can keep up with.

Intercom, the customer-messaging company, has reportedly 2–3× their PR throughput since leaning into AI-assisted development. That's a lot of code waiting in line. Their solution — and the inspiration for this tutorial — was to build an AI agent that scores the risk of each pull request and automatically approves the low-risk ones, so humans only spend time on the changes that genuinely need their judgment.

The claimed results are striking: AI-approved PRs are approved roughly five times faster than human-reviewed ones, and their revert rate (code that has to be rolled back after it ships) is reportedly lower with AI in the loop — not higher. That flips the usual assumption that removing human review means accepting more risk.

What Vercel Eve actually is (and why it matters for builders like you)

Eve is Vercel's framework for deploying AI agents — think of it as a pre-built scaffolding that lets you wire an AI brain to real tools like Slack and GitHub without writing a lot of plumbing code yourself.

What makes it relevant to designers and early builders specifically:

• It's directory-based. An Eve agent is configured through files in a folder, not through a complex dashboard or custom server setup. The structure is similar to other Vercel projects, so if you've deployed anything on Vercel before, the mental model carries over.
• Skills and tools are modular. You give the agent specific capabilities — like "read a pull request" or "post a comment" — as discrete add-ons, rather than writing one giant program.
• Vercel Connectors handle the annoying parts. Connecting your agent to Slack or GitHub used to mean wrestling with authentication tokens and refresh logic. Eve wraps all of that in a setup wizard. You click through it; it handles the technical handshake.

The underlying AI doing the actual code reading and risk assessment is powered by OpenAI's Codex model, which is trained specifically on code and can reason about what a change does — whether it touches critical paths, whether tests have passed, how wide the potential blast radius is if something goes wrong.

How a designer could actually build this

You don't need to be an engineer to follow the workflow Claire Vo demonstrates. Here's the practical shape of it:

1. Start a new Vercel project and add Eve. The Eve agent lives as a subfolder in your repo, with a configuration file that describes what the agent is and what it's allowed to do.

2. Define the agent's skills. For a PR review bot, the core skills are: read the diff (the list of changes in a pull request), score the risk based on criteria you define (blast radius, whether automated checks have passed, how much of the codebase is touched), and post a summary comment back to the PR.

3. Connect GitHub via Vercel Connectors. This is the wizard-driven step — no manual token management. Eve listens for new pull requests and triggers automatically.

4. Optionally connect Slack. The bot can post risk summaries to a Slack channel so the team sees what's been auto-approved and what needs eyes.

5. Set your approval threshold. You decide what "low risk" means. PRs below that threshold get approved automatically; the rest get flagged for human review.

The risk-scoring logic itself can draw on frameworks like "Diff Vader" (an open approach documented by Rewind) that breaks down factors like correctness signals and action-run status — but you can start simpler and refine over time.

The compliance question you'll probably have

If you're working with any kind of regulated data or enterprise client, your first instinct might be: there's no way I can auto-approve code without a human sign-off. It's a fair concern. The short answer, based on what Intercom has documented, is that it's possible within compliance frameworks like SOC 2 or HIPAA — but only if the approvals are logged, auditable, and explicitly covered in your code review policies. Eve's structured approach helps with the audit trail because every action the agent takes is traceable. That said: work with your security team before turning on auto-approval in a regulated environment. This isn't legal advice, and every situation is different.

The honest limits

Thirty minutes gets you a working bot, not a perfect one. You'll need to tune your risk thresholds based on your actual codebase — what counts as "low risk" for a marketing landing page is very different from what counts as low risk for a payment flow. And Eve is a Vercel-specific framework, so if your project lives elsewhere, this particular path doesn't apply (though the underlying concepts do).

The bigger shift here isn't the tool — it's the mindset. If you're building with AI, you're going to generate more change, faster, than old review processes were designed to handle. Building a lightweight review layer into your workflow early, before the queue becomes unmanageable, is the move. Eve makes that layer genuinely accessible to people who aren't full-time engineers. That's new, and it's worth experimenting with.