Agent view for every session

Use agent view to start and follow a small Claude Code task from one terminal dashboard. You will learn which sessions appear, how to answer a blocked session, and how to review a result before starting parallel work.

Before you start

• Claude Code CLI installed and authenticated on your computer, with access to a folder you trust.
• A terminal open in the website repository and one small read-only task, such as locating the contact form and its validation.
• A current CLI version with agent view enabled. This feature is a research preview; managed settings can disable it.

Want to do it yourself?

Difficulty

4 out of 10

Estimated completion time

20–40 minutes

Skills you’ll need: Identify the intended session, follow its visible progress, and verify the reported result against the actual files.

I’ll do it myselfI’d rather get help

How this helps your brand

Following the right session helps you answer blockers and review website investigations before starting more work. Checking the cited files keeps a completed status tied to an actual, useful result.

1. 1

   Step 1: Start in the right repository

   Check the folder shown by your terminal before opening Claude Code. For a first experiment, choose a read-only investigation so an unexpected result is easy to inspect.

   Run claude agents from your shell. If Claude asks whether you trust the workspace, read the folder path and continue only for the intended project. The dashboard lists background sessions; ordinary interactive sessions in other terminals are not automatically included.

2. 2

   Step 2: Dispatch one specific task

   Type the prompt below in the input at the bottom and press Enter. A row appears for the new background session. Give it a few moments to begin, then read its state and latest activity.

   Every new prompt entered in the dashboard starts another session. To clarify the task you just launched, open that row instead of putting a follow-up into the dispatch input.

   Illustrated exampleOne row to continue, one input to dispatch

   ● ● ●Agent view · simplified example

   Workspacenorthline-website

   • 1
     Locate contact form validationRead only · inspecting ContactForm.tsx
     Running

   2

   Dispatch input

   › Your next new task goes here…

   1 Continue this task

   Open its existing row. Ask “Which fields are required?” inside that conversation.

   2 Start another task

   A prompt in the dispatch input creates another background session.

   An ordinary interactive session in another terminal does not appear here automatically.

3. 3

   Step 3: Open the row when it needs you

   Use the arrow keys to select the session and press Space to inspect its peek panel. Press Enter on the selected row for the full conversation. Read the question or permission request and answer only what that task needs.

   After replying, press the left arrow on an empty prompt to return to the dashboard. The background session can continue without a terminal attached. These sessions run on your computer, so do not expect them to keep working through a shutdown.

4. 4

   Step 4: Verify the investigation result

   When the row reports a result, open its conversation and compare it with the brief. For the contact form example, check that the named file exists, the described fields match the page, and missing checks are clearly marked.

   Ask for a correction inside that session if a claim lacks evidence. A completed row is a useful signal to review the output, not a reason to accept every conclusion.

   Illustrated exampleTurn a completed report into checked evidence

   Completed report · example

   “The contact form requires an email address. The code submits to /api/contact.”

   A completed row is the start of your review.

   Compare with the actual file

   src/components/ContactForm.tsx

   <form action="/api/contact">
     <input
       name="email"
       type="email"
       required
     />
   </form>

   1. ✓
      File evidenceOpen the named file and confirm this code exists.
   2. ✓
      Supported claimThe email input has required. Its field name matches the report.
   3. ?
      Browser check still neededThis excerpt does not prove that a message is delivered. Test the real form separately.

   If the evidence is missing, ask for a correction inside this same session.

5. 5

   Step 5: Bring an existing session into the list

   For a regular Claude Code session you already use in another terminal, press the left arrow on an empty prompt to move it into the background and open agent view. Read any backgrounding dialog before continuing; some in-flight work may need to stop or restart.

   Begin parallel work only after you understand one session. Choose independent tasks, name the scope of each, and avoid having several sessions edit the same files at once.

6. 6

   Step 6: Recover without launching duplicate work

   If claude agents prints configured subagents and exits, the docs recommend updating Claude Code, then checking whether agent view is disabled by settings or policy. An empty dashboard is normal before you launch a background session.

   After a shutdown, inspect the stopped or failed row and its saved conversation before asking it to continue. Do not dispatch the same task again until you know whether the earlier attempt produced changes or results.

Inspect [repository or current project] to find how [specific website feature] works. Read files only. Do not edit, commit, publish, or send messages. Report the relevant file paths, the behavior supported by those files, and any gaps that need a browser check. Explain the result for a beginner and suggest one small next check. If the project or feature is ambiguous, ask me before broadening the task.