Build and validate a web app with browser tools
AI agents: See llms.txt for documentation discovery and navigation. Documentation articles are also available as Markdown by appending .md to the article URL. Join VS Code at Microsoft Ignite or catch the sessions online. A year of VS Code, wrapped. Explore the highlights. Build and validate a web app with browser tools View as Markdown Ask in VS Code Ask in VS Code Insiders Browser tools let an agent check a running web app instead of relying only on its source code.
The agent can open the app, interact with it, inspect the result, and repeat a failing interaction after a fix. In this guide, you build a calculator, validate it against explicit acceptance criteria, and practice reproducing and fixing a defect. Use the Copilot harness in the Chat view. The generated implementation can vary but the expected behavior stays the same. These interactive checks provide feedback during development.
They don't automatically create a saved test suite or prove that the entire app is correct. Prerequisites Install Visual Studio Code . Set up GitHub Copilot . Leave workbench.browser.enableChatTools Open in VS Code Open in VS Code Insiders turned on, which is the default. The sample uses HTML, CSS, and JavaScript without a build step or additional dependencies. Define the acceptance criteria Decide what the agent should observe before asking it to validate the app.
For this calculator, use the following checks: Check Actions Expected result Digit buttons Clear the calculator, then select a digit. Repeat for each digit from 0 to 9. The display shows the selected digit. Addition Enter 3 + 5 = . 8 Subtraction Enter 9 - 4 = . 5 Multiplication Enter 6 * 7 = . 42 Division Enter 8 / 2 = . 4 Clear Enter 9 + , select Clear , then enter 2 + 3 = . Clear displays 0 and resets the pending operation.
The new calculation returns 5 . Division by zero Enter 7 / 0 = . Error , not Infinity or NaN . Start each independent check with Clear so that a previous calculation doesn't affect the result. Step 1: Check browser tool availability Create a project folder and open it in VS Code. Open the Chat view ( ⌃⌘I (Windows, Linux Ctrl+Alt+I ) ), start a new chat, and select Copilot as the Session Target and Agent from the agent picker.
Select Open Customizations in the Chat view title bar > Tools and verify that Integrated Browser is selected. If the tools are missing, check workbench.browser.enableChatTools Open in VS Code Open in VS Code Insiders . Organization policy can also turn them off. See browser tools troubleshooting . Step 2: Build the calculator Ask the agent to build the app with the behavior you intend to verify: Create a calculator using index.html, styles.css, and script.js.
Use plain HTML, CSS, and JavaScript without dependencies or a build step. It must work when index.html is opened directly in a browser. Include buttons for digits 0-9, addition, subtraction, multiplication, division, Clear, and equals. Show the entered value and calculation result. Clear must reset the display to 0 and discard any pending operation. Division by zero must display Error rather than Infinity or NaN. Review the saved changes .
If edits are pending review, select Keep to retain them or Undo to revert them. Keep clears pending review; the files are already saved to your workspace. Don't assume the generated function names or internal structure match another calculator implementation. Judge the app against the acceptance criteria. Step 3: Run the acceptance checks Ask the agent to interact with the app without changing it during the check: Open index.html in the integrated browser and check the calculator by selecting its buttons.
Start each independent check with Clear. Verify: * Each digit button from 0 to 9 displays that digit after Clear. * 3 + 5 = 8, 9 - 4 = 5, 6 * 7 = 42, and 8 / 2 = 4. * Entering 9 + and then Clear displays 0. Entering 2 + 3 = afterward returns 5, without retaining the previous operation. * 7 / 0 = displays Error, not Infinity or NaN. Do not edit files. Report each check's actions, expected result, actual result, and pass or fail status.
Mark checks you couldn't run as unverified. Review the browser tool calls and reported results. A screenshot can support a visual observation, but it doesn't show that every interaction passed. If the agent can't open the page or complete an action, resolve that problem before treating the check as verified. The following video illustrates browser interaction with a calculator. Your generated layout and tool sequence might differ.
Step 4: Reproduce and fix a defect If an acceptance check failed, use that failure for this step. If all checks passed, introduce a controlled defect to practice the workflow: Introduce only this defect in the calculator: dividing by zero displays Infinity instead of Error. Keep the JavaScript syntactically valid and leave all other behavior unchanged. Do not fix the defect yet. Review the diff to confirm that only the intended behavior changed.
Then reproduce the defect before requesting a fix: Reload the calculator to use the saved code. Select Clear, enter 7 / 0 =, and report the actual display value. It should be Error. Do not edit files during this check. After you observe the incorrect result, ask for a focused fix: Fix the division-by-zero behavior so the display shows Error. Avoid unrelated changes. Reload the page and repeat the failing interaction. Then rerun every acceptance check from the earlier request, starting each with Clear.
Report the actual results and any checks you couldn't complete. For a different defect, replace the division-by-zero case with the actions and expected result from the failed check. Review the fix and verification output. Reload after code changes so the agent checks the updated app, not a page still running the previous code. Keep regression coverage Browser interaction during a chat isn't a saved automated test. To catch the same defect after future changes, add regression coverage to your project's test suite.
Use unit tests for isolated calculation behavior and browser tests for behavior that depends on the UI. Reuse an existing framework rather than introducing a second runner: Add a regression test for the defect we fixed using the project's existing test framework and conventions. Verify the expected behavior, not the implementation details, and run the relevant tests. If no suitable test framework exists, propose a minimal setup and wait for approval before installing dependencies or changing configuration.
Review the assertions and execution output. See Test existing code with AI for guidance on test quality, failures, and coverage. Apply the workflow to your web app For an existing project, give the agent the development-server command, app URL, a focused user journey, and expected results. Specify whether it should only report failures or also fix them. If a flow needs an existing browser session, share a page with the agent .
Sharing gives access to that page's current session, including its sign-in state. Use a test account and disposable data for actions that submit forms or change stored data. Other useful checks include: Forms : test valid and invalid input, error messages, and successful submission. Responsive layouts : compare navigation and content at specified desktop and mobile viewport sizes. Accessibility : investigate labels, headings, and keyboard interactions.
Agent checks aren't a complete accessibility audit; include manual keyboard and assistive-technology testing. For browser capabilities, session isolation, and privacy controls, see Use browser tools with agents . Next steps Use test-driven development to define behavior before implementation. Learn how to attach browser elements, screenshots, and console output to chat . 11/18/2026