Coding agents already consume compiler errors, test failures, and lint output. Interface work needs another signal: what the changed product actually looks like.
An agent-native visual loop should expose the flow, context, baseline, candidate, diff measurement, and review state through a stable machine-readable interface. The agent can run the flow, inspect a failed context, change the code, and publish a replacement candidate.
Keep approval separate from correction
Agents can diagnose and self-correct, but they should not silently redefine the approved baseline on protected branches. The merge check remains pending until a human accepts the visible result.
That separation produces a useful division of labor: machines iterate quickly on evidence; people decide whether the evidence is acceptable.
The result is not “AI replaces visual review.” It is a shorter path from a visual regression to a corrected candidate before scarce human attention is spent.