Agent View

The agent view, opened by the claude agents command, is a single control screen that lets you put several Claude Code sessions on one table, watch how they are doing, and step in only in the sessions that need a hand. Toss a bug fix, a PR review, and a flaky-test investigation each onto its own row, do other work, and come back only when a row is waiting for input or has produced a result.

Background reference

This page is background material on Claude Code itself, the platform MoAI-ADK runs on. MoAI-ADK’s own features are covered in the sections above it in the sidebar.

Info

One-line summary: Instead of scrolling transcripts one by one, see every running, waiting, and completed background session in one table and step in only at the moments that matter.

The agent view is an interface for managing background sessions — sessions that keep running without being tied to one terminal — from one screen. Each background session is a complete Claude Code conversation in its own right, and a separate supervisor process keeps it running even if you close the terminal. So you can toss a bug fix, a PR review, and a flaky-test investigation each onto its own row, go do other work, and come back when a row is waiting for input or has produced a result.

The agent view is in research preview and works on Claude Code v2.1.139 or later. Check your version with claude --version. The interface and shortcuts may change as the feature evolves.

Claude Code has several means of running work in parallel — the agent view occupies the “dispatch several unrelated full sessions and collect only results” seat among them.

The agent view exists because background execution became the default in Claude Code v2.1.198. The runtime puts sessions that do not need their results right away into the background and lifts them to the foreground only at the moment the result is needed. The agent view is precisely the control window for watching those background sessions in one place.

When a background session hits a tool that needs permission (e.g., Bash, WebFetch), the prompt is displayed on the main session screen. Since v2.1.186 the prompt even carries the name of which subagent is asking, and Esc can deny just that one call. Pre-listing frequently-used tools in the settings.json allow list before starting a long background task cuts the prompt frequency substantially.

Opening the agent view takes over the terminal and lists every session grouped by status. Sessions waiting for input and pinned sessions rise to the top, and each row shows the session name, current activity, and time since last change.

Needs input
  ✻ power-up design     needs input: double jump or wall climb?     1m

Working
  ✽ collision detection Edit src/physics/CollisionSystem.ts          2m
  ✢ playtest level 3    run 12 · all checkpoints cleared          in 4m

Completed
  ✻ title screen        result: menu, options, and credits done      9m
  ∙ sound effects       result: 14 SFX exported to assets/audio       4h

The icon at the front of each row indicates the session status through color and animation.

Separately, the icon shape indicates whether the underlying process is alive. ✻ (or animated ✽) means the process is alive and responds instantly; ∙ means the process has exited (you can still peek, reply, or attach, and Claude resumes from where it stopped); ✢ means a /loop session is waiting between iterations.

Each row’s one-line summary is generated by a Haiku-family model, so you know what a session is doing, wanting, and producing without opening the transcript. Working sessions refresh their summary at most every 15 seconds and at the end of each turn.

When a session opens a PR, a label like PR #1234 appears at the right end of the row, hyperlinked in terminals that support links. The PR number is colored by state.

For most work, this column is where you collect results — when a PR number turns green, review and merge. You can also dispatch a shell command as a background task by prefixing the input with !, like ! pytest -x; in that case no model is invoked, only the command runs on one row, and the latest output line is shown as its status.

Subagents or agent-team teammates spawned by a session are not listed as separate rows. Their output and progress merge into the parent session’s row summary and output. To see the details, peek into or attach to that session and view the full conversation.

For reference, since v2.1.219 subagent nesting is re-enabled by default, so one background session can fan out nested spawns up to depth 3 (disable with CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1). The agent view does not expand this nested structure row by row — it shows status by the outermost session.

The agent view shines when you have several mutually independent tasks Claude can advance without you watching every step.

• Monitoring long-running work: toss a long task like a flaky-test investigation onto a row, work in another window, and return when the row flips to needs-input or a result. Background sessions keep running even if you close the terminal or shell, thanks to the supervisor process.
• Tracking parallel work: dispatch a bug fix, a PR review, and a test investigation as three rows at once and compare states at a glance. File edits are isolated per session in worktrees under .claude/worktrees/, so all rows read the same checkout but each writes separately.
• Managing multiple projects on one screen: by default, background sessions from all projects appear in one table. To narrow to one project, specify the directory, e.g. claude agents --cwd ~/projects/my-app.

Each session consumes subscription usage independently. Running 10 agents in parallel burns your quota roughly 10x faster, so keep usage limits in mind before dispatching many at once. Parallelization is a trade — buying time by paying tokens. The reason MoAI-ADK prioritizes parallel fan-out for read-only investigation and review while running write work sequentially takes this cost structure into account alongside safety.

The basic flow cycles dispatch → observe → peek and reply → attach.

flowchart TD
    A["Run claude agents
Open the agent view"] --> B["Type a prompt + Enter
Dispatch a background session"]
    B --> C["Observe states in the table
Working / Needs input / Completed"]
    C --> D["Space
Peek panel + reply"]
    C --> E["Enter or →
Attach to the full conversation"]
    E --> F["← (empty input)
Detach and return to the table"]
    F --> C

A new background session starts via three routes.

# 1) Open the agent view, type a prompt in the bottom input, press Enter
claude agents

# 2) Start directly in the background from the shell
claude --bg "investigate the flaky SettingsChangeDetector test"

# 3) Designate a specific subagent as the main agent
claude --agent code-reviewer --bg "address review comments on PR 1234"

Every prompt entered in the agent view input starts a new session (it does not append to an existing one). To send an ongoing conversation to the background, run /background or its alias /bg inside the session, or press ← on an empty input.

When you designate a subagent definition as the main agent like the third route, you can fix the model per session via that agent file’s model field (the default inherit simply uses the main session’s model). For cost-sensitive collection work, set a cheaper model in advance; for reasoning-heavy work, set a higher-tier model — the model will not drift at dispatch time.

Attaching never stops a session. To end a session completely, run /stop inside it.

Press ? to view all shortcuts on screen. The most-used ones:

Warning

Worktrees Claude created for a session are removed along with it when deleting via double Ctrl+X, and uncommitted changes are lost. Push or commit first to preserve them.

You can also work directly with short IDs without opening the agent view.

claude agents --json        # print live sessions as a JSON array
claude attach <id>          # attach to a session in this terminal
claude logs <id>            # show a session's recent output
claude stop <id>            # stop a session
claude respawn <id>         # restart a session keeping the conversation

To fully disable the agent view and background agents, set disableAgentView to true or set the CLAUDE_CODE_DISABLE_AGENT_VIEW environment variable. You can put the setting in settings.json.

{
  "worktree": {
    "bgIsolation": "none"
  }
}

Setting worktree.bgIsolation above to "none" makes background sessions edit the working copy directly instead of moving into worktrees (v2.1.143+).

• Subagents
• Agent Teams
• Worktrees

• Manage multiple agents with agent view (Claude Code Docs)

Tip

To keep a long-running session responsive, pin it with Ctrl+T. An unpinned session left untouched for about an hour after finishing has its process stopped by the supervisor to reclaim resources, and it wakes a beat late on reattach.