Anthropic Agent View Claude Code Parallel Sessions Sandbox PoC

Purpose

Verify the current Claude Code Agent View surface safely before writing the article. The goal was to confirm the installed CLI version, inspect background-session commands, avoid uncontrolled account spend, and model the operational queue pattern with local-only session metadata.

Web Sources Checked

• Anthropic announcement: https://claude.com/blog/agent-view-in-claude-code
• Claude Code docs, parallel agents overview: https://code.claude.com/docs/en/agents
• Claude Code docs, Agent View: https://code.claude.com/docs/en/agent-view
• ClaudeUpdates v2.1.139 release tracker: https://www.claudeupdates.dev/version/2.1.139
• Simon Willison Code w/ Claude live blog context: https://simonwillison.net/2026/May/6/code-w-claude-2026/

Commands And Outputs

Installed CLI

command -v claude
claude --version

Output:

/opt/homebrew/bin/claude
2.1.139 (Claude Code)

CLI Help Surface

claude --help
claude agents --help

Relevant output:

Usage: claude [options] [command] [prompt]

Commands:
  agents [options]                                  Manage background and configured agents

Usage: claude agents [options]

Manage background and configured agents

Options:
  -h, --help                   Display help for command
  --setting-sources <sources>  Comma-separated list of setting sources to load
                               (user, project, local).

Background Session Smoke Check

The command below was intended as a help check, but Claude Code treated it as a background-session start. The session was immediately stopped.

claude --bg --help
claude stop 817cc05f

Output:

Starting background service...
backgrounded · 817cc05f (idle — send a prompt to start)
  claude agents             list sessions
  claude attach 817cc05f    open in this terminal
  claude logs 817cc05f      show recent output
  claude stop 817cc05f      stop this session

stopped 817cc05f

This confirms the installed CLI accepts --bg and creates a shell-manageable background session ID. It does not prove a full autonomous task was completed.

Agent List Command In Non-Interactive Shell

claude agents

Output excerpt:

36 active agents

Plugin agents:
  accessibility-compliance:ui-visual-validator · sonnet
  agent-orchestration:context-manager · inherit
  ...

Built-in agents:
  claude · inherit
  Explore · haiku
  general-purpose · inherit
  Plan · inherit
  statusline-setup · sonnet

In this non-interactive capture, claude agents printed configured agent types rather than a full-screen TUI session roster. The official docs describe the interactive Agent View table as a terminal UI.

npm Package Version Check

npm view @anthropic-ai/claude-code version dist-tags --json

Output:

{
  "version": "2.1.143",
  "dist-tags": {
    "stable": "2.1.133",
    "latest": "2.1.143",
    "next": "2.1.143"
  }
}

Local CLI was 2.1.139, while npm reported 2.1.143 as latest on 2026-05-16.

Local Queue Model

This local-only Node.js prototype models the dashboard behavior developers need operationally: sort blocked sessions first, keep running/idle/done visible, and surface the next session needing intervention.

rm -rf /tmp/effloow-agent-view-poc-2026-05-16
mkdir -p /tmp/effloow-agent-view-poc-2026-05-16
cd /tmp/effloow-agent-view-poc-2026-05-16
node - <<'NODE'
const sessions = [
  { id: 'a11ce001', title: 'review auth boundary', state: 'blocked', last: 'needs approval for shell command', repo: 'api' },
  { id: 'b00b1350', title: 'write migration tests', state: 'running', last: 'executing focused test suite', repo: 'billing' },
  { id: 'c0ffee42', title: 'summarize flaky spec logs', state: 'done', last: 'summary ready', repo: 'web' },
  { id: 'd15patch', title: 'prototype dashboard copy', state: 'idle', last: 'waiting for prompt', repo: 'ui' },
];
const priority = { blocked: 0, running: 1, idle: 2, done: 3 };
const rows = sessions
  .sort((a, b) => priority[a.state] - priority[b.state] || a.repo.localeCompare(b.repo))
  .map(s => `${s.state.padEnd(8)} ${s.id} ${s.repo.padEnd(8)} ${s.title} :: ${s.last}`);
const counts = sessions.reduce((acc, s) => (acc[s.state] = (acc[s.state] || 0) + 1, acc), {});
console.log('Agent view queue summary');
console.log(JSON.stringify(counts, null, 2));
console.log(rows.join('\n'));
console.log('next_action=' + (sessions.find(s => s.state === 'blocked')?.id ?? 'none'));
NODE

Output:

Agent view queue summary
{
  "blocked": 1,
  "running": 1,
  "idle": 1,
  "done": 1
}
blocked  a11ce001 api      review auth boundary :: needs approval for shell command
running  b00b1350 billing  write migration tests :: executing focused test suite
idle     d15patch ui       prototype dashboard copy :: waiting for prompt
done     c0ffee42 web      summarize flaky spec logs :: summary ready
next_action=a11ce001

What Worked

• Claude Code was installed locally at /opt/homebrew/bin/claude.
• The installed version was 2.1.139, matching the Agent View release line.
• claude --bg was accepted and created a manageable idle background session with attach, logs, and stop commands.
• The idle background session was stopped successfully.
• npm package metadata was reachable and showed newer 2.1.143 as latest on 2026-05-16.
• A local-only queue model reproduced the operational reason Agent View matters: blocked sessions need to rise above running, idle, and completed work.

What Failed Or Was Not Attempted

• A live autonomous background task was not launched because it could spend account quota and require user-specific permissions.
• /bg was not tested inside an interactive Claude Code session because this automated run is non-interactive.
• The full-screen Agent View TUI was not captured; non-interactive claude agents output exposed configured agent types instead.
• No claims are made about throughput, benchmark improvements, or production reliability.

Limitations

This PoC verifies CLI availability, version, background-session creation/stop behavior, npm package freshness, and a local queue-management model. It does not verify completion of a real Claude Code coding task, PR creation, worktree merge behavior under live edits, or organization-level managed settings.