Claude Code agent view is a terminal screen that shows all your Claude Code background sessions in one place: which ones are working, which are waiting for your answer and which are done. You open it with claude agents or with the left arrow key on an empty prompt. Every prompt you type there starts its own session in the background. Press Space to peek into a session and reply, Enter to attach to it fully, and the left arrow to return to the list. The sessions keep running through a background service even when you close the terminal. So you don't need tmux or a wall of terminal tabs for this. Before a session started this way edits files, Claude moves it into its own git worktree so parallel sessions don't get in each other's way. Agent view has been available as a research preview since May 2026, and every session uses your quota on its own.
My own server workforce runs each employee in a tmux session with Remote Control, as described in Claude Code on a Server: tmux, Remote Control. This article explains agent view based on Anthropic's docs as of October 26, 2026, and sorts out where it replaces tmux and where it doesn't. Everything about Claude is collected on my Claude page.
What agent view is, and what it isn't
According to the docs, Claude Code has five ways to work on several tasks at once. They differ in who coordinates and where the work runs:
So agent view is the approach where you hand out the work yourself: give away several independent tasks, see at a glance what is going on, and only step in when a session needs you. According to the docs, these sessions report their results only to you; they only message each other if you use cross-session messaging. Subagents a session starts don't appear as separate rows. Watch the names: /agents inside a session is not the same as claude agents in your shell.
How subagents work inside a session is covered in Claude Code Subagents: Best Practices; how to split work across several agents in general, in Orchestrating Multiple AI Agents.
Agent view or tmux?
Anthropic describes the problem agent view is meant to solve in its announcement of May 11, 2026: anyone running agents in parallel used to juggle multiple terminal tabs, a tmux grid and an overloaded mental ledger. Agent view replaces exactly that part.
On your own computer, then, you no longer need tmux to run several Claude sessions in parallel. A remote server is a different matter: for Remote Control there, the docs explicitly recommend tmux or screen so the session survives the end of the SSH connection. Whether agent view's background service survives logging out on a server is not described in the docs, and I have not tested it. My server workforce still runs in tmux.
One detail if you combine the two: a session you attach to from agent view always renders in fullscreen mode. According to the docs, tmux copy mode then shows only the visible viewport; to scroll back, use Page Up and Page Down or Ctrl+O.
Quick start in five steps
- Open it: run
claude agents in your shell. If you haven't trusted the folder yet, that dialog comes first.
- Hand off a task: type a prompt at the bottom and press Enter. A new background session starts and appears as a row. Every further prompt starts another session; it does not go to the previous one.
- Peek and reply: select a row with the arrow keys and press Space. The preview shows the latest output or the question the session is waiting on. Type a reply and press Enter.
- Attach and detach: Enter or the right arrow opens the full session. The left arrow on an empty prompt takes you back to the list. Detaching never stops the session.
- Bring in an existing session: type
/bg in a regular session, or press the left arrow on an empty prompt. The session moves to the background and appears in the list.
If you want claude with no arguments to always open agent view, turn on "Open agents view by default" in /config, or set it directly with /config defaultToAgentsView=true.
Reading the list
The list groups sessions by state, with pinned sessions and those that need you at the top. The icon at the start of each row shows the state:
When a session opens a pull request, a number such as #1234 appears at the right edge of the row. Its color shows the status: yellow while checks or review are pending or checks have failed, green when checks have passed and no review is blocking, purple once merged. Ctrl+S groups the list by folder instead of state, and Ctrl+T pins a session to the top.
According to the docs, the short summary in each row is written by a small Haiku-class model, at the end of every turn and every few minutes during long turns. Each of these is one short request through your normal provider, billed like the session itself. In between, the row updates from the session's own output without a request.
Start and manage sessions from the shell
You don't have to open agent view to start a background session:
claude --bg "investigate the flaky SettingsChangeDetector test"
Use --name to give the session a name of your own, and --agent to run one of your subagents as the main agent:
claude --bg --name "flaky-test-fix" "investigate the flaky SettingsChangeDetector test"
claude --agent code-reviewer --bg "address review comments on PR 1234"
The prompt is a regular argument, not a -p value; Claude Code rejects --bg combined with -p. For a plain shell command without Claude there is claude --bg --exec 'pytest -x'; Claude Code keeps its output only in memory and cleans it up about five minutes after the command exits.
After the start, Claude Code prints a short ID and the commands to manage the session. The most important ones according to the docs:
Since version 2.1.290, claude attach and claude logs also accept part of a name instead of the ID. If you want to read the state from outside, for example for your own status display, the docs say to query claude agents --json rather than the files under ~/.claude/jobs/. How to customize Claude Code's status bar yourself is covered in Claude Code Statusline: Setup and Examples.
Worktrees: how parallel sessions stay out of each other's way
A session you start from agent view or with claude --bg begins in your working directory. Before it edits files, Claude moves it into its own git worktree under .claude/worktrees/. All sessions read the same checkout, but each writes to its own copy. This does not apply to a session you send to the background yourself with /bg or the left arrow: it keeps working where it was. Outside a git repository there is no isolation, so don't set parallel sessions loose on the same files there.
The docs are clear about what happens to the work:
- Commit and push: Claude commits in its worktree without asking and pushes the branch if there is a remote.
- Draft pull request: Claude opens one when the task calls for it.
- Never: pushing to
main or master, force-pushing, merging.
- Your rules take precedence: if the task, your
CLAUDE.md or memory says you handle commits and pushes yourself, Claude leaves git to you.
Be careful when deleting: if you remove a session in agent view by pressing Ctrl+X twice, Claude Code also deletes the worktree including uncommitted changes, according to the docs. If the worktree has commits that aren't saved anywhere else, Claude Code refuses on the first attempt and names the branch. For repositories where worktrees are impractical, turn isolation off with "worktree": { "bgIsolation": "none" } in .claude/settings.json.
Model, effort and permissions
What you type into agent view's input is usually a new task. A few commands act on the list itself, though: /model, /effort and /fast set what the next sessions start with. /effort and /fast have been available there since version 2.1.299. You set fixed defaults for every session from one window when you open it:
claude agents --permission-mode plan --model opus --effort high
A session you send to the background with /bg keeps its permission mode. A newly started session begins the way a new claude session in that folder would, unless you set defaults when opening agent view. For bypassPermissions mode, Claude Code asks for your explicit acceptance here too, because a session nobody is watching then acts without approval. For unattended sessions, that is a risk you take on deliberately, not a convenience switch.
The most important keys
If you type a filter into the input, the list narrows: s:blocked shows everything waiting on you, n:login searches names and first prompts. The name filter n: (version 2.1.287) and search with Ctrl+F (2.1.288) arrived in early October 2026. Agent view itself is older: Anthropic introduced it on May 11, 2026, with version 2.1.139.
Limits
- Quota: background sessions use your subscription quota just like interactive ones. According to the docs, ten sessions in parallel use it roughly ten times as fast as one.
- Local only: the sessions run on your machine. They survive sleep but not a shutdown; their saved history can be picked up again afterward.
- Background service: if the service stops, sessions on macOS and Linux end after about a minute unless you start
claude agents.
- Research preview: according to Anthropic, the interface and keys may still change.
- Turning it off:
disableAgentView: true or the CLAUDE_CODE_DISABLE_AGENT_VIEW environment variable turns off background sessions and agent view; admins can enforce this through managed settings.
If your tasks don't need your computer at all and live in a GitHub repository, there is also the route through cloud sessions, which keep running when your machine is off. How they work is covered in Claude Code on the Web: Cloud Sessions.
Frequently asked questions
What is Claude Code agent view?
A terminal screen that shows all your Claude Code background sessions in one place, with their state, latest message and age. You start new sessions there, answer questions and attach to a session when needed.
How do I open agent view?
With claude agents in your shell, or with the left arrow on an empty prompt in a running session. With /config defaultToAgentsView=true, claude with no arguments always opens agent view.
Do I still need tmux with agent view?
Not on your own computer to run several Claude sessions in parallel: a background service keeps the sessions running even when the terminal is closed. On a remote server over SSH, the docs still recommend tmux or screen for Remote Control.
What is the difference between agent view and subagents?
Subagents are helpers inside one session that Claude starts itself and whose result comes back into your session. Agent view shows complete, independent sessions that you start yourself and that report only to you.
Does agent view cost extra?
No, but every background session uses your quota like a regular session. On top of that, there is one short request per session to a small model for the row summary at the end of each turn.
Do sessions keep running when I close the terminal?
Yes. A separate background service runs them, so you can close agent view, the shell and the terminal. When the machine shuts down they stop; their history stays saved and you can pick them up again afterward.
Where to go from here
Start with two small, independent tasks in a git repository and watch how Claude moves them into their own worktrees and reports back at the end. Write in your CLAUDE.md whether Claude may commit and push on its own before you run more sessions in parallel. If you want to run Claude Code permanently on a server, read Claude Code on a Server: tmux, Remote Control; for steering from your phone, Claude Code Remote Control from Your Phone. How others organize their parallel sessions and AI employees is what we discuss in my German-speaking community Claude Practitioners.
