Should an AI agent update an existing pull request?

Should I let my AI agent update an existing pull request instead of opening a new one

Yes, when the agent is continuing the same change on the same branch. GitHub pull requests are designed to evolve as feedback arrives, and reviewers expect new commits, updated diffs, and fresh checks on the same pull request while the work is still the same change. GitHub’s docs also describe branches, commits, and pull requests as the normal way to keep work isolated and reviewable.

Use one pull request when the new agent work still answers the same question: fix the failing tests, adjust the same feature, respond to review comments, or clean up the same refactor. Updating the existing pull request keeps the conversation, approvals, checks, and file history in one place, which is exactly how GitHub expects review to work.

Open a new pull request when the agent has crossed into a different change. That includes a different feature, a different bug, a different base branch, or a follow-up that needs a different reviewer set and a different decision. GitHub explicitly supports stacked pull requests for dependent changes, which is the better pattern when one change must land before the next one can be reviewed cleanly.

The part people get wrong is thinking “more commits” means “new pull request.” It does not. A pull request is a review thread around a branch, and GitHub lets you keep pushing new commits to that branch while the pull request stays open. A new pull request only helps when the old one would make the review harder to understand.

The inconvenient part is that an AI agent can blur the boundary between “same change” and “new change” fast. If the agent starts fixing unrelated lint issues, small cleanup tasks, or extra behavior that no reviewer asked for, the original pull request becomes noisy. At that point, splitting the work into a new branch and a new pull request usually saves time, because reviewers can approve the original diff without re-reading everything.

A good rule is simple: if the new commit would still belong in the original pull request description, keep updating it. If you would need a new title, a new summary, or a new set of acceptance criteria, open a new pull request. GitHub’s review flow is built around clear, focused pull requests, and it is easier for humans to approve a narrow diff than a mixed one.

For AI agents, updating an existing pull request is usually the safer default because it preserves review context. Reviewers can see what changed after feedback, GitHub can rerun checks on the same branch, and the thread stays tied to one history of decisions. That matters when the agent is acting as a helper on your own repository, not as a way to scatter many near-duplicate pull requests.

Use a new pull request when the existing one already has final feedback and the next step is a separate piece of work. A clean split is better than overloading one thread with multiple goals. If the first pull request is about changing an API and the second is about updating the docs, the docs change usually deserves its own pull request unless the reviewer explicitly asked for both together.

If the agent needs to keep the branch current, GitHub supports updating the pull request branch from the pull request page by merging or rebasing the base branch into the head branch. That keeps the branch in sync without forcing a new pull request just because the target branch moved. Rebase is useful when you want a linear history, merge is useful when you want the branch to reflect the base branch update plainly.

One practical guardrail helps a lot: let the agent update an existing pull request only when it can name the exact reason for the change in one sentence. If the agent cannot say whether it is fixing review feedback, resolving merge drift, or completing the original task, the branch is probably being used for more than one job. That is when a new pull request keeps the work honest and readable.

The best default is this: same task, same branch, same pull request. Different task, different branch, different pull request. That keeps review history useful, keeps checks attached to the right change, and avoids making reviewers guess which commits belong to which decision. GitHub’s own documentation is structured around that workflow, not around reopening a fresh pull request for every extra commit.

If you want a lightweight place to track the agent’s work and coordinate reviews, DevConnect keeps the workflow centered on the task instead of the noise around it: https://devconnectplatform.com. Use that kind of structure to keep the branch purpose clear, then let the pull request stay open until the change is actually complete.

A final test is whether the new work would confuse code owners. If yes, split it. If no, keep the existing pull request open and push the update. That choice protects approvals, keeps the diff legible, and avoids turning one review into several half-finished ones.