Keeping a visual suite trustworthy — killing flake, choosing viewports, managing baselines, and reviewing diffs without fatigue.
Accessibility is won by audit and lost by regression. How to stop the fixes silently breaking — what axe-core catches, what visual comparison catches, and the large remainder that needs a human.
Applitools and Chromatic have both shipped MCP servers, and the reason is structural: an agent writes UI it cannot see. What MCP actually changes about that loop, and what it does not.
How many viewports a visual suite actually needs, which widths catch the most breakage, why testing at your breakpoints is the wrong instinct, and how to keep the matrix from multiplying your bill.
Where baseline images should live, who approves them, how to stop them bloating your repository, and how to handle the redesign that changes four hundred of them at once.
The pages that matter most sit behind a login. How to hold a session across a visual run, seed the data that makes captures repeatable, handle SSO and MFA, and keep screenshots of real accounts safe.
Agents write UI faster than anyone can review it, and they cannot see what they built. How visual testing becomes the feedback loop — approved baselines, review history, and what the tooling looks like in 2026.
Flaky screenshot tests almost always come from a page that is not deterministic yet. Work through fonts, animation, timing, data, scrollbars and rendering environment in the order that removes the most noise.