Sign in Subscribe
news

Claude Code Agent View

Sebastien Dubois

Sebastien Dubois

— 4 min read

Canonical version: Claude Code Agent View.

Agent view, opened with claude agents, is one screen for dispatching and managing many background Claude Code sessions. Each row shows what a session is doing, what it needs, or what it produced — without scrolling through transcripts. Requires v2.1.139+, the release that introduced it on 2026-05-11 (the same one that shipped /goal); in research preview, so the UI and shortcuts may change.

Background sessions are full Claude Code conversations hosted by a per-user supervisor process, so they keep running with no terminal attached. Close agent view or the shell and they keep going. They survive sleep (the supervisor reconnects on wake) but stop on shutdown or reboot: within 48 hours they show as failed and restart from where they left off when you attach or reply; past 48 hours they show as stopped and you press Enter twice to resume.

Why it matters

Replaces the "many terminals, many tmux panes, many half-forgotten sessions" pattern with one inbox-style view. Dispatch a bug fix, a PR review, and a flaky-test investigation as three rows; check back when one says it needs input or has a result.

Quick start

  • claude agents — opens the view. Esc to leave; sessions keep running
  • Type a prompt + Enter — dispatches a new background session as a new row
  • Space on a row — peek panel (last output or pending question); type a reply and Enter to send
  • Enter / → on a row — attach (full interactive session, as if you had run claude)
  • ← on an empty prompt — detach back to the table (also works from any session to background it)
  • /bg inside a running session — sends it to the background and into agent view

Session states

Icon color/animation shows state; icon shape shows whether the process is alive.

  • Working (animated), Needs input (yellow), Idle (dimmed), Completed (green), Failed (red), Stopped (grey)
  • ✻ / animated ✽ — process alive; ∙ — process exited (still resumable); ✢ — /loop session sleeping between iterations
  • After ~1h idle with no terminal attached, the supervisor kills the process. Peek/attach restarts it from where it left off.

Pull-request status

When a session opens a PR, a coloured dot appears on the right (yellow = waiting/failed checks, green = passed, purple = merged, grey = draft/closed). For most tasks this row is where the work gets picked up: dispatch → green dot → review → merge.

Dispatch patterns

  • <agent-name> <prompt> or @<agent-name> — runs that custom subagent as the session's main agent
  • @<repo> — runs the session in a sibling repo under the directory agent view was opened from
  • /<skill> — suggests skills to dispatch as the prompt (recurring tasks → package as a skill)
  • #<number> or a PR URL — selects the existing session for that PR instead of starting a new one
  • Shift+Enter — dispatch and immediately attach
  • Ctrl+G — open the prompt in $EDITOR for long inputs
  • Filter the list (no dispatch) with a:<name>, s:<state> (incl. s:blocked), or #<number> / PR URL

Worktree isolation

Before editing files, a background session moves itself into an isolated Git Worktree under .claude/worktrees/. Parallel sessions read the same checkout but each writes to its own. Merge or push before deleting the session — the worktree is removed with it. Outside a git repo, no isolation: avoid parallel edits to the same files.

Defaults that travel

Flags passed to claude agents apply to itself and every session it dispatches:

  • --permission-mode <mode> — bypassPermissions / auto require having accepted them once interactively first (since a session you aren't watching would act without approval)
  • --model <name>, --effort <level>
  • --settings, --add-dir, --plugin-dir, --mcp-config, --strict-mcp-config
  • --cwd <path> scopes the view to one project (incl. its worktrees)

Shell management

Every session has a short ID, printed when started via claude --bg "<prompt>" (combine with --agent <name> and --name <display>).

  • claude attach <id> — open in this terminal
  • claude logs <id> — recent output
  • claude stop <id> / claude kill <id>
  • claude respawn <id> / claude respawn --all
  • claude rm <id> — remove from list (cleans worktree if no uncommitted changes)

Limitations

  • Local-only — sessions run on your machine; preserved across sleep, stopped by a shutdown
  • Quota — each session consumes subscription usage independently; 10 in parallel ≈ 10× the burn rate
  • Worktree destruction on delete — push/merge first
  • No nested sessions — subagents and teammates a session spawns don't appear as rows
  • Disable entirely via disableAgentView setting or CLAUDE_CODE_DISABLE_AGENT_VIEW

How it differs from siblings

  • vs AI Subagents: subagents run inside one session and only report back; agent view sessions are independent processes with their own conversation
  • vs Claude Code Agent Teams: teams coordinate via shared mailbox + task list and can message each other; agent view sessions are siblings, not collaborators
  • vs Claude Code Web: web runs sessions in Anthropic's cloud and survives your machine sleeping; agent view runs locally
  • vs Claude Code Routines: a routine starts a cloud session by itself on a schedule, an API call or a GitHub event, with your machine off. Agent view sessions run on your machine and you dispatch each one. Local scheduled work (/loop) does show up here, as a ✢ row sleeping between iterations
  • See Run agents in parallel for the full comparison

References

About Sébastien

Ready to get to the next level?

Found this valuable? Share it with someone who needs it.

Join 6,000+ readers. Get practical systems for knowledge & AI. Free.

Subscribe ✨

Free: Knowledge System Checklist

A clear roadmap to building your own knowledge system. Subscribe and get it straight to your inbox.

6,000+ readers. No spam. Unsubscribe anytime.

Subscribe