Choose one visitor journey
Test a concrete path: open the landing page, choose a guide, switch language and follow the CTA. Write down what the visitor should see. Start locally or on staging. A page loading successfully does not prove its buttons, forms or mobile layout work.
Limit browser access
Chrome DevTools MCP appeared in the radar and exposes browser tools to coding agents. A connector can reveal browser data to the model. Use a separate test profile without personal sessions. Review source, installation and telemetry settings under your policy before using a new connector.
Have the agent actually click
Ask the agent to execute the route and record what happened at each step. Use available browser tools for snapshots, screenshots and console or network errors. Name untested behavior. A screenshot shows appearance; pair it with the action and URL to support a behavioral claim.
Save a reproducible failure
A useful bug report includes the URL, viewport, steps, expected behavior and observed behavior. Add relevant evidence. Do not let the agent state a cause as fact when it only observed a symptom. A reproducible failure is a better repair starting point than “the website is broken”.
Repeat the failed step
After a change, rerun the original steps and compare with the acceptance criterion. Check the adjacent steps if the change could affect them. Save the retest with the report. Mark a flow passed only when the expected behavior is observed.
Your worksheet
# Browser check Test URL: [your test or staging website] Initial state: [logged out / clean test session] Visitor goal: [one goal, such as opening a guide] | Step | Action | Expected | Actual | Evidence | | 1 | Open [URL] | [starting page] | [finding] | [screenshot] | | 2 | Click [exact button] | [destination] | [finding] | [screenshot] | | 3 | Check [element] | [content or state] | [finding] | [screenshot] | Desktop: [viewport and result] Mobile: [viewport and result] Bug: [reproduction steps + expected/actual] Fix: [change] Same-route retest: [pass/fail and evidence] No real purchases, messages or customer data in this trial.
STARTER
Make it concrete.
Test this allowed staging URL using the separate test profile. Follow my four steps on desktop and mobile. Record expected and observed behavior, URL and viewport. Save relevant screenshots and console errors. Use test data only, send nothing and make no purchases. Mark untested behavior and create a reproducible report for each failure.
Sources and further reading
The AI Operator · C.11, M3.D.1, M3.D.4, M3.D.6, M2.5
github.com/ChromeDevTools/chrome-devtools-mcp
github.com/ChromeDevTools/chrome-devtools-mcp/blob/main/docs/tool-reference.md