Anthropic shipped Agent View as a Research Preview in Claude Code v2.1.1391 on May 11, 2026, and the most consequential change is not any single feature but the architectural assumption it overturns. Until now, running multiple Claude Code sessions in parallel meant tmux panes, stacked terminal windows, or shell background jobs with no visibility into what each agent was doing. Agent View replaces that with a full-terminal dashboard that lists every session, surfaces its state, and lets you peek or reply without attaching to the full transcript.
What Agent View Is and Why It Changes the Workflow
The claude agents command opens a dashboard that groups sessions by state2: Working, Needs input, Idle, Completed, Failed, or Stopped. Each row displays the session name, a state icon, a one-line activity summary, and how long ago it last changed. The summaries are generated by a Haiku-class model and refresh at most once every 15 seconds2. This is the kind of interface developers have been assembling themselves with terminal multiplexers and custom scripts. Anthropic is now shipping it as a first-class TUI.
The shift from serial to parallel agent orchestration matters because it lowers the friction of fanning out a plan-execute pattern across several concurrent sessions. You can kick off a background session to refactor a module while another traces a bug and a third drafts tests, then check on all three without losing context in any of them.
The Dashboard: Four Signals and Keyboard Navigation
The dashboard surfaces four signals per row: the session name, a state icon that encodes the group, a one-line activity summary, and how long ago the session last changed. Pressing Spacebar on a selected row opens a peek panel2 that shows the session’s most recent output or the specific question it is blocking on. You can reply inline from the peek without attaching to the full transcript, which matters when you just need to unstick an agent, not read its entire thought process.
Enter or the right arrow key attaches2 to the selected session’s full transcript. The left arrow detaches back to Agent View. Detaching never stops the session. The keyboard navigation is intuitive enough that you can move between monitoring and intervention without reaching for the mouse.
Dispatching and Backgrounding: /bg, —bg, and Attach/Detach
There are two ways to get a session into Agent View. The /bg command2 (also /background) moves your current interactive session into the background. The claude --bg "[task]" flag2 launches a fresh background job directly from the shell, bypassing the interactive chat entirely. This is the pattern you will use for fire-and-forget tasks: lint fixes, dependency updates, test generation.
Worktree Isolation and the Supervisor Process
Background sessions isolate file edits using git worktrees under .claude/worktrees/2. This prevents collision when multiple agents touch the same repository. The worktree is deleted when the session is removed, which means uncommitted changes are lost unless you merge or push them before cleaning up. Treat each background session as a disposable branch with a built-in expiration date.
A per-user supervisor process hosts background sessions locally. The architecture is sensible for a CLI tool: no cloud state to sync, no remote session persistence to secure. The tradeoff is fragility. Sleep your MacBook and the sessions stop. Leave them idle for an hour and they get killed to free resources. This is agent orchestration for the workstation, not the server.
The /goal Command: Autonomous Loops in the Same Release
Agent View was not the only addition in v2.1.139. The /goal command3, shipped in the same release, sets a completion condition and lets Claude work across multiple turns until the condition is met. An overlay shows elapsed time, turn count, and token usage. The combination of /goal and /bg suggests Anthropic is building toward a model where you describe an outcome, background the attempt, and check Agent View later to see whether it finished, failed, or got stuck waiting for clarification.
Quota Economics and the Parallel-Session Cost Trap
The economics are linear and unforgiving. There is no batch discount for concurrent sessions, no special background rate. If you spin up six agents to explore different approaches to a refactor, you pay for six streams of reasoning, six sets of tool calls, six context windows. Agent View removes the technical barrier to parallelism but leaves the financial barrier intact. For teams, this will show up in the monthly invoice before it shows up in the sprint velocity.
Free-Plan Lockout and the Pro Floor
Agent View is available on Pro, Max, Team, Enterprise, and Claude API plans1. The Free plan is not supported. This is worth stating clearly, but it is not a new paywall. Claude Code itself requires a Pro subscription or higher1; Agent View inherits that floor rather than adding to it. The practical effect is the same: a solo developer whose workflow goes parallel must be on Pro. The distinction matters only for framing. Anthropic is not gating backgrounding behind an extra surcharge. They are gating Claude Code behind Pro, and Agent View comes with it.
Research Preview Caveats and Known Issues
Anthropic labels Agent View a Research Preview, which means the UI and keybindings may change without warning1. The docs explicitly warn against hard-coding output parsing into automation scripts. This is not a stable API surface. It is a user interface that happens to be rendered in the terminal.
There is already at least one known bug. A GitHub issue filed May 12, 20264 reports that sleeping or looping background agents are incorrectly shown in the “Ready for Review” section instead of an active state. The misclassification is cosmetic but confusing: an agent burning quota in an infinite loop looks finished until you attach and scroll. Expect more edge cases as the preview matures.
Frequently Asked Questions
Does any competing coding agent offer a comparable multi-session dashboard?
No. Cursor and Windsurf expose one agent conversation at a time through tab or panel UIs, and GitHub Copilot has no background-session concept at all. Developers on those platforms who want parallel agent workflows must still cobble together tmux layouts or custom orchestration scripts. Agent View is the first native TUI that treats concurrent sessions as a first-class monitoring surface.
Do the dashboard row summaries themselves cost anything?
Yes. Each summary is a Haiku-class inference call billed against the user’s plan. With a refresh ceiling of once per 15 seconds per session, a dashboard left open with six background agents generates roughly 24 paid requests per minute even if the user never interacts. That hidden cost compounds on top of the primary model’s token usage for each session.
Can Agent View manage sessions on a remote machine over SSH?
Not reliably. The per-user supervisor process is local and filesystem-bound to git worktrees under .claude/worktrees/. Sessions terminate when the host sleeps, shuts down, or the SSH connection drops. There is no session persistence or reattachment protocol, so a network hiccup mid-refactor can leave a worktree with stale uncommitted changes that will be deleted on cleanup.
What’s the practical risk from the state-misclassification bug?
An agent spinning in an infinite loop or stuck in a sleep cycle appears under “Ready for Review” (issue #58381), which means it continues consuming quota while looking finished. Until the fix ships, the safe pattern is to attach briefly to any completed agent and scroll its transcript before queuing dependent work — the dashboard state alone cannot be trusted as a reliable signal during the Research Preview.