How to Build a 24‑7 Autonomous AI Agent Team with OpenClaw

Background

The author runs Unwind AI and the Awesome LLM Apps repository, handling six daily tasks (research, tweets, LinkedIn posts, newsletters, GitHub reviews, issue triage) that each take 30–60 minutes. A single monolithic agent proved ineffective, so six specialized AI agents were hired.

Team members

Monica (Chief Operating Officer) : coordinates the team, makes strategic decisions, and assigns work.

Dwight (Research) : scans X, Hacker News, GitHub trends, AI blogs, and papers three times a day, producing structured intelligence.

Kelly (X/Twitter) : consumes Dwight’s research and drafts tweets in the author’s voice.

Rachel (LinkedIn) : uses the same research to craft LinkedIn posts with a leadership angle.

Ross (Engineering) : handles code review, bug fixes, and technical implementation.

Pam (Newsletter) : turns Dwight’s daily intel into newsletter excerpts.

Hardware and installation

All components run on a Mac Mini M4, though any always‑on computer (laptop, gaming PC, $5/month VPS) works. OpenClaw supports macOS, Linux, and Windows (WSL). Installation requires two terminal commands:

# 1. Install OpenClaw
curl -fsSL https://openclaw.ai/install.sh | bash

# 2. Quick start
openclaw onboard

The command starts the gateway, background processes, and keeps agents running after the terminal closes.

File structure

workspace/
├── SOUL.md               # Monica (root agent)
├── AGENTS.md             # behavior rules for all agents
├── MEMORY.md             # Monica’s long‑term memory
├── HEARTBEAT.md          # self‑healing cron monitor
├── agents/
│   ├── dwight/
│   │   ├── SOUL.md
│   │   ├── AGENTS.md
│   │   └── memory/
│   ├── kelly/
│   │   ├── SOUL.md
│   │   ├── AGENTS.md
│   │   └── memory/
│   ├── ross/
│   │   ├── SOUL.md
│   │   └── memory/
│   ├── rachel/
│   │   └── ...
│   └── pam/
│       └── ...
├── intel/
│   ├── DAILY-INTEL.md   # Dwight’s research output
│   └── data/
│       └── 2026-02-11.json

Monica runs in the root directory; other agents are sub‑agents that she can invoke or that run on independent cron schedules.

SOUL.md files

Each agent is defined by a SOUL.md file that specifies identity, role, principles, and relationships. Example for Dwight:

# SOUL.md (Dwight)

## 核心身份

**Dwight** — 研究脑。以 Dwight Schrute 命名，因为你跟他一样严厉：彻底到变态、领域里啥都知道、把工作极其当回事。不废话。不瞎猜。只给事实和来源。

## 你的角色

你是团队的情报骨干。研究、验证、组织、传递情报给别人创建内容。

**你供给：**
- Kelly（X/Twitter）— 病毒趋势、热推文、突发新闻
- Rachel（LinkedIn）— 思考领导力角度、行业新闻

## 你的原则

### 1. 永远别编造
- 每个声明带来源链接
- 每个数据来自源头，不估算
- 不确定就标 [未验证]
- "我不知道" 比瞎说强

### 2. 信号 > 噪音
- 不是所有趋势都重要
- 优先：与 AI/agents 相关性、互动速度、来源可信度

Monica’s SOUL.md similarly defines strategic oversight, task delegation, and an “no‑fluff” operating style.

Collaboration

Agents do not call APIs, use message queues, or rely on orchestration frameworks; they simply read and write markdown files. Dwight writes intel/DAILY-INTEL.md, Kelly and Rachel read it to draft content, and Pam creates the newsletter. Coordination is purely file‑system based, which avoids API rate limits and authentication issues.

Scheduling and self‑healing

OpenClaw’s built‑in cron runs each agent on a fixed schedule. The author’s actual schedule (illustrated below) runs Dwight first, followed by Kelly and Rachel, ensuring the research output exists before content generation.

Heartbeat runs a health check after each cron cycle. If a job’s lastRunAtMs is stale (>26 hours), the heartbeat forces a rerun via: openclaw cron run <jobId> --force Example job IDs:

- Dwight morning (8:01 AM): 01f2e5c5-3a83-4018-a725-dee59e54733e
- Kelly viral (9:01 AM, 1:01 PM): c9458766-78bb-4eeb-b8f4-d63dc1f0e601
- Ross engineering (10:01 AM): b12b2fc6-dd7d-4123-b904-2148a5cfb70b
- Dwight afternoon (4:01 PM): 19ff40e4-b1b0-4d32-9d24-753ac2cf8f46
- Kelly X draft (5:01 PM): 05da0c81-39e1-4d06-bdcd-2dfab4562ba4
- Rachel LinkedIn (5:01 PM): 9819bc6b-7e36-406f-b0c3-d80ca383d914

Heartbeat also monitors cron failures, machine restarts, and network interruptions, automatically re‑triggering missed jobs.

Interaction

There is no web UI; the author interacts with agents via Telegram, which OpenClaw can expose as a bot. Monica handles most conversations, delegating to other agents when needed. A typical morning: the author opens Telegram, sees Dwight’s research summary, reviews Kelly’s three tweet drafts and Rachel’s LinkedIn post, approves them, and proceeds with coffee—all in about ten minutes.

Optimization

Agents start with a rough prompt (“corrective prompt engineering”) and improve over weeks. Feedback such as “no emojis, no tags” updates Kelly’s memory file, making future drafts correct on the first try. Dwight’s principle was refined to prioritize signal over noise, reducing irrelevant trends. Naming agents after TV characters provides an instant persona baseline that is later refined through iterative memory updates.

Security

Each agent runs on its own scoped API keys and isolated mailbox on the Mac Mini, never accessing the author’s personal accounts. Services like Gemini, Eleven Labs, and Claude have separate keys that can be revoked instantly if anomalous usage is detected.

Troubleshooting

Gateway crashes are rare; restart with openclaw gateway restart.

Stale cron jobs are auto‑re‑run by the heartbeat.

Context window overflow is mitigated by keeping SOUL.md under 60 lines and loading only today’s and yesterday’s memory files.

Output quality drops when memory files become contradictory; heartbeat periodically consolidates daily logs into a clean MEMORY.md.

File‑write conflicts are avoided by a “one‑write‑many‑read” pattern (e.g., only Dwight writes DAILY-INTEL.md).

The overarching lesson: start with a single agent and a single task, achieve reliability, then incrementally add more agents.

Cost

Hardware: Mac Mini M4 starts at $499, but any always‑on machine works (old laptop, $5/month VPS). Model costs (monthly): Claude Opus $200, Gemini API $50‑70, TinyFish $≈ 50, Eleven Labs $≈ 50, Telegram free, OpenClaw open‑source. Total under $400/month for a fully autonomous team, saving 4–5 hours of manual work each day.

Actionable recommendations

Week 1: Deploy a single agent (e.g., research) on OpenClaw, set up Telegram, and monitor a cron job.

Week 2: Add memory optimization and iterative prompt tweaks.

Week 3: Introduce a second agent (e.g., content drafting) and configure file‑based handoff.

Week 4+ : Continue adding agents only when a clear workflow gap is identified.

Conclusion

After a month the agents feel like a continuously working team rather than an on‑demand tool. The real moat lies not in the underlying LLMs but in the surrounding system—SOUL.md definitions, memory files, scheduling, and coordination mechanisms—that turn raw models into a reliable, self‑improving workflow.