Agent-written code increases review throughput, but it also changes the economics of attention. Humans cannot manually open every affected route and reconstruct every permission, theme, and locale before merge.
Put the evidence in the pull request
Choose critical flows, define their contexts as code, and run them against the PR head SHA. A useful visual check groups results by flow and context, shows the previous approved state beside the candidate, and links every decision back to the commit.
Keep the status pending while any result is unreviewed. Pass only when every change is approved. Fail immediately when a reviewer rejects a result.
Let agents use the same feedback
Structured visual results can also feed the coding loop. An agent can learn that the German navigation overflowed or the member role lost its empty state, make a correction, and publish a new candidate. Human approval remains the boundary that decides what ships.
Code review explains how the implementation changed. Visual review explains what the user will experience. AI-velocity teams need both.