Turn One Computer into a Fleet of Autonomous AI Agents with Claude Code /loop

What is Claude Code /loop?

Claude Code is Anthropic’s AI programming agent that can execute tasks autonomously. The newly added /loop command turns Claude from a simple Q&A tool into a background agent that repeatedly runs a user‑defined task at a specified interval.

One‑line usage

Simply type a command such as:

/loop 5m 检查部署是否完成，如果失败了告诉我原因

Claude will execute the task every five minutes, reporting any failures, while you continue working on other things.

Supported time units

The interval can be expressed in seconds ( s), minutes ( m), hours ( h) or days ( d), using standard cron‑style syntax. The task runs only when Claude is idle, so it never interrupts an active conversation.

Real‑world examples

/loop 5m 检查部署是否完成

– monitor deployment status every five minutes.

/loop 20m 去 GitHub 看看有没有新的 issue，有的话分类打标签

– automatically triage new issues.

/loop 1h 跑一遍测试套件，把失败的用例列出来并尝试修复

– run the test suite hourly and attempt fixes.

/loop 30m 分析 logs 目录下最近的错误日志，生成摘要

– generate a log‑analysis summary every half hour. /loop 检查构建状态 – default 10‑minute interval when no unit is given.

Natural‑language reminders

下午 3 点提醒我推送 release 分支
45 分钟后，检查集成测试是否通过了

A single session can host up to 50 concurrent loops, which automatically expire after three days.

Nesting loops

You can embed another command inside a loop, e.g.:

/loop 20m /review-pr
/loop 1h /simplify

This allows “Skill → Loop” and “Loop → Skill” compositions.

Building a complete autonomous agent

The loop engine alone is not enough; you need two additional components:

1. CLAUDE.md – the agent’s brain

Place a CLAUDE.md file at the project root describing conventions, coding standards, and common workflows. Claude reads this file on startup to understand the project context.

# 项目：*** 监控系统
## 核心规则
- 所有 API 改动必须同步更新 openapi.yaml
- 测试覆盖率不低于 80%
- commit message 遵循 conventional commits
## 常见操作
- 部署命令：`make deploy-**`
- 日志路径：`/***/logs/`
- 告警检查：`curl -s http://**:9090/api/v1/alerts`
## /loop 任务清单
- 每 30 分钟检查一次告警接口
- 每 2 小时跑一次 lint + test
- 每天早上 9 点生成昨日变更摘要

2. Git hook for persistence

Configure a PostToolUse hook in .claude/settings.json to automatically commit changes after each loop execution:

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Bash|Write|Edit",
        "command": "cd /*** && git add -A && git diff --cached --quiet || git commit -m 'auto: agent checkpoint' --no-verify"
      }
    ]
  }
}

This approach saves tokens, guarantees no data loss, and makes every change traceable via git log. You can also define a logging convention in CLAUDE.md to record each loop’s result.

3. Multi‑terminal deployment

Open separate terminals for different projects and run independent /loop commands, e.g.:

# Terminal 1 – Frontend
cd /***/project-a && claude --dangerously-skip-permissions
> /loop 10m 跑一遍 npm test，失败的用例尝试修复

# Terminal 2 – Backend
cd /***/project-b && claude --dangerously-skip-permissions
> /loop 30m 检查 GitHub 新 issue 与 PR，自动处理

# Terminal 3 – Data
cd /***/project-c && claude --dangerously-skip-permissions
> /loop 1h 读取用户行为数据，生成异常指标报告

In theory, a single machine can host unlimited agents as long as the API quota permits.

Pricing (brief)

Claude Pro costs $20 / month with standard /loop limits. Claude Max 5x is $100 / month (five times the Pro quota) and Claude Max 20x is $200 / month (effectively unlimited, recommended for multiple agents).

Frequently Asked Questions

Q: Does a /loop task stop when the terminal is closed? A: Yes. The task follows the session lifecycle and expires after three days. For persistent scheduling, Claude Code Desktop offers scheduled tasks independent of the terminal.

Q: Will /loop interrupt an ongoing conversation? A: No. Tasks run only when Claude is idle.

Q: Is the --dangerously-skip-permissions flag safe? A: It bypasses permission prompts, suitable for controlled automation. For critical files, define explicit permission boundaries in CLAUDE.md or omit the flag.

Q: How does Claude Code compare to OpenClaw? A: OpenClaw excels at cross‑platform notifications and 24/7 operation. Claude Code offers tighter integration with the development environment—direct file access, command execution, and Git interaction—making it a more natural fit for daily coding workflows.

Conclusion

By combining the /loop engine, a well‑crafted CLAUDE.md brain, and Git‑hook persistence, developers can create robust, self‑maintaining AI agents that handle repetitive DevOps tasks, freeing human time for higher‑level problem solving.