Give browser agents the full story.

Whether an agent navigates the browser or helps an engineer investigate a recording, the handoff should answer a few concrete questions. Keep it focused on the task.

01The task

Say what to investigate and what a correct result would look like.

Find why a profile name reverts after saving.

02The starting state

Include the route, environment, account role, and setup that matter.

Test workspace · /settings · member account.

03The target

Identify the control and its surrounding section. Add a locator or relevant markup when available.

The Save changes button inside Profile.

04The sequence

List the actions in order, including the transition where the problem appears.

Edit the name → save → reload.

05The observation

Separate what you saw from what you think caused it.

“Profile saved” appears, but the old name returns.

06The evidence

Link the reviewed recording or screenshot, note when it was captured, and say what is still unverified.

Recording at 00:12 · persistence not yet investigated.

A screenshot shows appearance. DOM and accessibility information add structure and names. A recording shows the captured sequence. Choose the evidence that answers the question, and say when something is missing.

A profile name reverts after a reload. The page has two “Save changes” buttons. Add context below to see what each version explains—and what the receiving agent still needs to find out.

The name did not save. See the attached settings screenshot.

Investigate the name field in Profile on /settings.
The target is the Save changes button in that section.

page.getByRole('region', { name: 'Profile', exact: true })
  .getByRole('button', { name: 'Save changes', exact: true })

Task: investigate a profile name that reverts after saving.
Environment: test workspace, /settings, member account.
Start: Profile name is "Avery".
Target: Save changes inside the Profile section.

Steps:
1. Change the name to "Avery Chen".
2. Select Save changes in Profile.
3. Wait for "Profile saved", then reload the page.

Expected: the name is still "Avery Chen" after reload.
Observed: "Profile saved" appears, but "Avery" returns.
Evidence: attach the reviewed recording, its capture time,
and the reload moment (00:12 in this fictional example).

Scope: use the test account only; leave Preferences unchanged.
Report what you verify and what remains unknown.

An authored teaching example, not a live agent session. The sample screen and prompts contain fictional data.

In this example, the save message appears but the name does not survive a reload. Checking only the message would miss the reported bug.

import { test, expect } from '@playwright/test';

test('profile name survives a reload', async ({ page }) => {
  // Configure baseURL and an authorized test account first.
  await page.goto('/settings');
  const profile = page.getByRole('region', {
    name: 'Profile', exact: true,
  });
  const name = profile.getByRole('textbox', {
    name: 'Display name', exact: true,
  });
  const save = profile.getByRole('button', {
    name: 'Save changes', exact: true,
  });

  await expect(save).toHaveCount(1);
  await name.fill('Avery Chen');
  await save.click();
  await expect(profile.getByRole('status'))
    .toHaveText('Profile saved');

  await page.reload();
  await expect(name).toHaveValue('Avery Chen');
});

Playwright locators resolve against the current DOM. Its actionability checks help determine whether a control can be clicked. Keep a separate assertion for the application result.

Fill in the details that matter to the task. Put the relevant moment next to the evidence link so the receiving person or agent knows where to look.

Task: [What should the agent investigate?]
Environment: [Test URL, build, browser, account role]
Starting state: [Relevant setup and test data]
Target: [Section + role/name; locator or markup if available]

Steps:
1. [Action]
2. [Action]
3. [Transition that exposes the issue]

Expected: [Observable result, including after reload if relevant]
Observed: [What actually happened; keep guesses separate]
Evidence: [Reviewed link + capture time + relevant moment]
Access: [Confirm the intended recipient can open the evidence]
Allowed actions: [What can be inspected or changed, and where]
Still unknown: [Missing data or checks not yet performed]

Re-check the current page before acting.
Return: steps attempted, results observed, and remaining gaps.

Review screenshots, recordings, URLs, and markup before sharing. Leave out credentials and unrelated personal data. Use an authorized test environment for actions that change application data.

Model Context Protocol (MCP) gives compatible clients a shared way to access a server’s resources and capabilities. What the agent can read or do depends on that server and its permissions.

Samelogic’s MCP connection provides read-only access to authorized recordings, steps, and available element evidence for clients such as Claude Code and Codex. The agent can use that material while investigating; an engineer still reviews and verifies a proposed fix.