Agent Browser Validation

Date: 2026-03-24
Tool: agent-browser 0.22.0
Scope: verify whether newly installed agent-browser can extend current frontend/admin E2E closure beyond the existing Playwright CDP path

Validation Goal

Determine whether agent-browser can:

act as a stable browser automation path in the current constrained Windows environment
close any remaining OS-level validation gap
be promoted to a supported project E2E entrypoint

PowerShell wrapper entrypoint agent-browser.ps1 is blocked by execution policy in the current environment.
The native binary is callable directly:
- C:\Users\Admin\AppData\Roaming\npm\node_modules\agent-browser\bin\agent-browser-win32-x64.exe
agent-browser requires its daemon socket directory to be redirected to a writable root:
- AGENT_BROWSER_SOCKET_DIR=C:\Users\Admin\.codex\memories\agent-browser-home\socket
Official browser bootstrap path was also tested:
- agent-browser install
- result: failed to fetch Chrome for Testing version metadata from googlechromelabs.github.io
- conclusion: the official first-run install path is not currently closed in this environment

Tested with:

--executable-path C:\Program Files (x86)\Microsoft\Edge\Application\msedge.exe
--executable-path C:\Users\Admin\AppData\Local\ms-playwright\chromium-1208\chrome-win64\chrome.exe
crash-dialog / crashpad suppression args aligned with the existing stable browser wrapper:
- --noerrdialogs
- --disable-breakpad
- --disable-crash-reporter
- --disable-crashpad-for-testing
- --no-sandbox
- --disable-dev-shm-usage
- --headless=new

Result:

failed before DevTools became ready
representative error:
- Chrome exited early ... without writing DevToolsActivePort
- crashpad / mojo access-denied errors were still present
the failure reproduced both with system Edge and with Playwright cached Chromium chrome.exe

Conclusion:

agent-browser launch mode is not currently usable as a stable browser launcher in this environment

Tested by connecting agent-browser to the already stabilized external browser path started by frontend/admin/scripts/run-cdp-smoke.ps1.

Result:

stable for observation steps:
- open
- wait --load networkidle
- snapshot -i
- get text
successful snapshot captured interactive refs on /login, including:
- username textbox @e7
- password textbox @e8
- login button @e6
partially usable for limited in-page actions:
- focus @e7: succeeded
- eval against the page context: succeeded for DOM value injection
not stable for user-like interaction steps:
- fill timed out repeatedly after successful snapshot
- type @e7 e2e_admin: failed with read timeout / host not responding
- keyboard inserttext e2e_admin: failed with CDP error (Input.dispatchKeyEvent): Invalid 'text' parameter
- click @e6 after successful eval-based field population: failed with read timeout / host not responding
- eval-triggered DOM submit.click() returned successfully, but the subsequent wait --url **/dashboard still failed with read timeout / host not responding

Representative result:

open http://127.0.0.1:3000/login: success
snapshot -i: success
get text @e2: success
focus @e7: success
eval to set username: success
fill @e7 e2e_admin: failed with read timeout / host not responding
type @e7 e2e_admin: failed with read timeout / host not responding
keyboard inserttext e2e_admin: failed with invalid CDP parameter
click @e6: failed with read timeout / host not responding
DOM-triggered login + wait --url **/dashboard: failed with read timeout / host not responding

Conclusion:

in the current environment, agent-browser connect mode is usable for observational inspection and limited in-page diagnostics
it is not reliable for user-like interactive E2E execution or workflow progression

Installing agent-browser does not currently close the remaining OS-level validation gap.
It does not replace the existing supported E2E path.
Current supported browser-level E2E path remains:
- Playwright library
- external browser launch
- CDP connection
- npm.cmd run e2e:full:win
agent-browser is currently best described as:
- installed
- partially usable for read-only browser inspection plus limited DOM-assisted diagnostics
- not yet a supported interactive E2E runner
- not evidence of full OS-level automation closure
- not capable of supporting full simulated user-operation or OS-level validation in the current environment