Reviewing AI agent work
Published July 28, 2026 · by the AQ team
Reviewing AI-written code well means reviewing more than the diff: you want the session that produced it, the actual behavior of the running app, and a way to send corrections back to the agent instead of fixing its work by hand. In AQ, all three live in the workspace behind the pull request.
Agents raise review volume and change its failure modes. The classic agent bug is a change that reads clean and does something subtly different from what the PR description claims. A diff view cannot catch that; running the code catches it immediately. AQ is built so the reviewer always has the running code.
Open the workspace behind the PR
Every agent-opened PR is tracked on the workspace that produced it, and the workspace is the review context: the terminal scrollback shows what the agent tried, what failed, and what it did about it; the editor shows the working tree as it actually is. Reviewing "why did it write this" stops being archaeology, because the reasoning is in the scrollback two panes away.
Click the preview before you approve
The workspace's dev server runs as a live preview. The reviewer uses the feature: clicks the new flow, resizes the window, tries the edge case from the ticket. For agent code this is the single highest-value review step, and in AQ it costs one click instead of a local checkout, install, and run. When something is off, pin a comment on the preview at the exact spot, and send it to the agent as its next prompt. The loop is: review, comment, agent revises, re-check.
Bring in reviewers who do not read code
Half of "is this right" is a product question. Preview links work for people outside the team: a stakeholder opens the running app without an account, and a review link that allows commenting lets them sign in with their email and comment in place. Their feedback lands in the same channel as the engineer's, and either can be forwarded to the agent verbatim.
The PR still lands in GitHub
None of this replaces your review process. Agents open normal pull requests on your repos with your branch protections and your required approvals; AQ tracks each PR's state (opened, merged, closed) on its workspace so the team sees which agent work actually shipped. The complementary guide, written for the GitHub side of the flow, is reviewing pull requests from AI agents.
Frequently asked questions
What is different about reviewing AI-written code?
Volume and failure mode. Agents produce more PRs, and their characteristic bug is plausible-looking code that does not quite do what the description says. That failure mode is caught by running the code, which is why AQ keeps a live preview one click from every diff.
Can review feedback go straight back to the agent?
Yes. Comments pinned on the workspace preview can be sent to the agent as its next prompt, so a review note becomes the revision instruction without being retyped.
Does AQ bypass our branch protections?
No. Agents commit and open PRs through your normal GitHub flow, so branch protections, required reviews, and CI apply unchanged. AQ adds review context around that flow; it does not shortcut it.
Can someone outside the team review the work?
Yes, at the preview level: the workspace's running app has a shareable link that works without an account for viewing, and a comment-enabled review link (quick email sign-in) for pinning feedback. Workspace access itself stays owner-managed.