Can an AI Agent Open Pull Requests on Your Repo? How It Works

Yes — an AI agent can open Pull Requests on your repo through a GitHub App you install with scoped permissions. It creates a branch, commits a change, and opens a PR through the GitHub API. Crucially, a PR is a proposal: it changes nothing until you review and merge it. The agent suggests; you decide.

The mechanism, step by step

A PR-based agent works exactly like a careful human contributor who doesn't have merge rights:

1. You install a GitHub App on the specific repositories you choose, granting scoped permissions — usually read/write on contents (to create a branch and commit) and read/write on pull requests (to open one). It doesn't get your whole account; it gets those repos, those scopes.
2. The agent authenticates as the App, requesting a short-lived installation token rather than holding a long-lived personal credential.
3. It creates a branch, commits the change, and opens a PR against your default branch.
4. You review the diff and the description, then merge or close. Nothing in your codebase changes until you click merge.

The PR is the safety boundary

The reason "an AI opening PRs" is safe where "an AI editing your code" is scary is the Pull Request itself. A PR:

• Shows the exact diff before anything happens.
• Runs your CI — tests, type-checks, linters — on the proposed change.
• Can carry a preview deploy so you can click through the change before merging.
• Requires a human merge. The agent has no permission to push to a protected branch or merge its own PR.

So the worst case isn't "the agent broke production" — it's "the agent opened a PR you closed." That's a fundamentally different risk profile.

What opening a PR looks like in code

Under the hood it's ordinary GitHub API usage. A minimal example with Octokit:

import { Octokit } from '@octokit/rest'

async function openPr(octokit: Octokit, owner: string, repo: string) {
  // 1. branch off the default branch, 2. commit the change to it
  //    (create blob/tree/commit, update the new branch ref) …
  // 3. open the Pull Request
  await octokit.pulls.create({
    owner,
    repo,
    head: 'velyr/improve-hero-cta',
    base: 'main',
    title: 'Move the hero CTA above the fold',
    body: 'Most visitors never scroll to the current CTA. This moves it up.',
  })
}

There's no magic — it's a branch and a pulls.create. The intelligence is in what change to make; the mechanism for proposing it is standard and reviewable.

What it can't do

A well-designed PR agent deliberately can't:

• Push to your protected default branch.
• Merge its own Pull Request.
• Touch repositories you didn't select during install.

Those limits are the point — they keep a human in the loop on every change.

Why this model beats a dashboard

The alternative — a tool that tells you what to fix and leaves you to implement it — puts all the work back on you. A PR agent does the implementation and asks for a yes/no. You keep full control (you merge) while skipping the manual edit. That's the model Velyr uses: it finds your highest-impact conversion fix each week and opens it as a Pull Request for your approval.