Claude Code Agent View: The Multi-Session Dashboard Guide

Anthropic shipped Claude Code Agent View on May 11, 2026 — a built-in CLI dashboard that surfaces every active Claude Code session in a single keyboard-driven table. Before Agent View, anyone running parallel agents juggled five-to-eight terminal tabs or a hand-rolled tmux grid, with no way to see at a glance which session was waiting on input. Agent View collapses all of that into one screen.

This guide covers how to open Agent View, the three ways to launch new sessions, the peek panel that lets you reply without attaching, the new /goal command for long-running agents, keyboard shortcuts, how it stacks up against OpenAI Codex Desktop, and the rate-limit math you need to know before running ten sessions in parallel.

What Is Claude Code Agent View?

Agent View is a single CLI dashboard that lists every active Claude Code session in a keyboard-driven table. Each row shows the session's current state (Running, Waiting, or Done), the most recent output line, and when you last interacted with it. The interface is deliberately minimal: one table, one input field at the bottom, every active session visible without scrolling.

The launch is part of a broader shift Anthropic has been making toward treating Claude Code as a multi-agent development platform rather than a single coding assistant. The same launch window brought dreaming, multi-agent orchestration for up to twenty specialists, and the Managed Agents outcomes API. Agent View is the local control plane that ties all of those parallel workflows together for individual developers.

How to Open Agent View

There are two ways to launch the dashboard. From a fresh terminal, run claude agents. From inside any already-active Claude Code session, press the left arrow key to drop directly into Agent View without losing your place. Press Esc to exit at any point — sessions keep running in the background, and you can return whenever you like.

Agent View requires Claude Code v2.1.139 or later. Check your version with claude --version and update via your package manager if you're below the cutoff.

The fact that Esc returns to a normal terminal without killing sessions is the design detail that makes the feature actually useful. You can flick into the dashboard, peek at three running agents, dispatch a fourth, and pop back out — all without losing whatever you were doing in your primary session.

Understanding Session States

Every session in the table is in one of three states. Each state matters in a different way:

If you take one thing away from this guide, it's this: Waiting is the state you scan for. The whole point of multi-agent development is offloading work in parallel — and that only works if you can tell at a glance which agents have hit a question they need you to answer.

Three Ways to Launch a New Session

Agent View doesn't replace how you start sessions — it adds three dispatch methods, all of which keep the new session running without a terminal tab attached to it.

1. Type a task into the Agent View input. The input field at the bottom of the dashboard accepts any prompt. Press Enter, and the new session shows up as a row in the table immediately.
2. Background an active session with /bg. From inside any running session, the /bg slash command sends it to the background. It keeps working without holding a terminal tab open, and Agent View picks it up as a normal row.
3. Launch already-backgrounded from the shell. Run claude --bg "your task description" from any terminal. The session starts in the background and appears the next time you open Agent View.

All three methods are equivalent — the difference is just how you got there. In practice, most developers find themselves using /bg when they realize a session is taking longer than expected, and claude --bg when scripting parallel work from a build pipeline.

The Peek Panel — Replying Without Attaching

The peek panel is the workflow that turns Agent View from a nice feature into a genuine force multiplier. Select any session row in the dashboard, press Space, and a side panel opens showing that session's most recent output. From there you can interact without ever fully attaching to the session:

• Enter — type a reply inline and send it back to the session.
• Number keys (1-9) — when the session is asking a multiple-choice question, just press the option number.
• Tab — auto-fill a suggested reply, then edit and send.
• ! prefix — send a Bash command for the session to execute (useful for "run the tests" or "git status").
• → (right arrow) — fully attach to the session and see the entire transcript.
• ↑ / ↓ — move between session peeks without closing the panel.

The number-key shortcut is the unsung hero here. Most of the time a Waiting session is asking something like "Do you want me to (1) refactor in place, (2) add a wrapper, or (3) revert?" — and being able to clear that prompt with a single keystroke from the dashboard, without context-switching into the session, is what makes running five or six agents genuinely sustainable.

/goal: The Long-Running Agent Companion

The same launch window introduced /goal, a slash command designed for long-running autonomous work. Where a normal prompt is one task, /goal takes an outcome — for example, "build a working pull request for issue #123" or "refactor all auth modules to use the new SDK" — and the agent works toward that outcome across multiple sub-tasks autonomously.

Paired with Agent View, the value is obvious: you dispatch three or four /goal sessions in the morning, walk away, and check Agent View periodically. The dashboard shows you which goals are still grinding, which have hit a question they need you to resolve, and which have finished. For recurring or scheduled goals, the table also surfaces the next planned run time.

Keyboard Shortcuts Reference

Agent View is keyboard-first. Memorize these — there is no mouse-friendly path.

Agent View vs OpenAI Codex Multi-Agent

OpenAI's Codex Desktop also offers a multi-agent workflow, but the design philosophies are different in ways that matter. Codex runs cloud tasks on OpenAI's clone of your repo — fire-and-forget, review the diff when it's ready. Agent View runs locally on your filesystem, lets you steer mid-task, and never copies your code anywhere.

Neither is strictly better. Codex wins when you want to dispatch a batch of independent issues and come back hours later — its "review diffs in a queue" model is excellent for that. Agent View wins when you're actively coding and want to spin up help on adjacent problems without losing visibility into them.

Real-World Multi-Agent Use Cases

A few patterns where Agent View earns its keep:

• Parallel git worktrees. Dispatch one session per feature branch — each in its own worktree — and let Agent View be the cross-branch radar.
• Specialized agents. One agent does the implementation, another writes tests in parallel, a third reviews the diff when both are done. Agent View shows you when the reviewer hits Waiting so you can answer its clarifying question.
• Long /goal runs in the background. Kick off a refactor or migration goal and let it work for an hour. Glance at Agent View every fifteen minutes to clear any prompts.
• Bug triage swarm. Drop three or four reproduce-and-isolate sessions on different open issues. Each works independently; you only engage with the ones that hit Waiting.

Rate Limits and Token Cost Considerations

Running five parallel sessions burns rate limits and tokens roughly five times faster than one. The compute is independent — there is no shared pool that dedupes overlapping work. A few practical implications:

• Use Sonnet 4.6 for routine sessions and reserve Opus 4.7 for the ones that actually need the reasoning depth. The Advisor Strategy (route hard sub-problems to Opus on demand) keeps cost under control.
• Most developers find 4–8 parallel sessions is the sweet spot. Below four, you're not getting much leverage. Above eight, coordination overhead and rate limit throttling start eating the marginal benefit.
• If you're running heavy parallel workloads from a laptop, hosted environments matter. A VPS dedicated to Claude Code (such as Hostinger's Claude Code VPS plan) keeps sessions running even when you close your laptop and reduces local CPU strain when many sessions are active simultaneously.

Team and Enterprise admins who want to limit Agent View can disable it via the disableAgentView managed setting — useful if your org's billing model doesn't accommodate easily-multiplied parallel usage.