Building Your First AI Agent with OpenClaw

You've installed OpenClaw. The gateway is running. Now what? This tutorial takes you from an empty workspace to a fully functional personal AI agent — with personality, memory, channel connections, and the foundation for automation.

We'll build this step by step, testing at each stage. By the end, you'll have an agent that knows who it is, remembers your conversations, and is available in your messaging apps.

Step 1: Understand the Workspace

Your agent lives at ~/.openclaw/workspace/. If you installed via Homebrew, this directory already exists with some defaults. If it doesn't, create it:

The workspace will contain these files by the time we're done:

Step 2: Write Your SOUL.md

This is the most important file. It defines your agent's personality, communication style, and behavioral rules. (For a deep dive, see our complete SOUL.md guide.)

Start with something practical:

This is a starting point. You'll refine it over the first two weeks as you notice behaviors you want to adjust.

Step 3: Create Your USER.md

USER.md gives your agent context about you:

Be specific. "Works in tech" is useless. "Senior DevOps engineer managing 3 AWS clusters for a fintech company" gives context that shapes every response.

Step 4: Set Up AGENTS.md

AGENTS.md is the operating manual — how your agent handles memory, tasks, safety, and workflow:

Step 5: Initialize Memory

Seed MEMORY.md with starting context:

This file will grow organically. In a week, it'll contain useful context. In a month, it'll be invaluable. See our Memory System guide for advanced strategies.

Step 6: Connect Your First Channel

Telegram is the easiest first channel:

1. Open Telegram and message @BotFather
2. Send /newbot and follow the prompts
3. Copy the bot token BotFather gives you
4. Add it to OpenClaw:

Find your bot in Telegram and send it a message. If everything is configured, your agent responds. That first response — from an AI running on YOUR hardware, with YOUR personality configuration — is a genuinely exciting moment.

Step 7: Test and Iterate

With the basics in place, test with these prompts:

1. "Tell me about yourself" — Should reflect your SOUL.md
2. "What do you know about me?" — Should reflect USER.md and MEMORY.md
3. "What's 2+2?" — Should be direct, not verbose
4. "I'm feeling stressed about a deadline" — Tests emotional intelligence
5. "Search the web for [something current]" — Tests skill access

Pay attention to what feels off. Too verbose? Add a rule to SOUL.md. Too formal? Adjust communication style. Doesn't remember something? Check MEMORY.md.

Step 8: Add a Second Channel

Once Telegram works, add another channel:

Multi-channel is where the agent experience truly clicks. Same personality, same memory, different interface. See our WhatsApp guide for details.

Step 9: Enable Basic Automation

A reactive agent is useful. A proactive agent is transformative. Create HEARTBEAT.md:

For a full deep dive on automation, see our Automation guide.

Step 10: Grow Over Time

Your agent isn't done — it never is. The best agents evolve continuously:

• Week 1: Refine SOUL.md based on interactions. Fix personality quirks.
• Week 2: Memory accumulates. Conversations feel more personal.
• Week 3: Add more skills. Explore capabilities.
• Month 2: Automation becomes central. Morning briefings, monitoring, proactive alerts.
• Month 3+: The agent becomes indispensable.

The turning point usually happens around week 3 — when memory accumulation and automation combine to create an experience qualitatively different from any chatbot.

Common First-Agent Mistakes

Over-Engineering SOUL.md on Day 1

Don't spend 3 hours on SOUL.md before talking to your agent. Write a basic version, start chatting, iterate. Real interactions teach more than theory.

Not Seeding Memory

An agent with empty MEMORY.md gives generic responses. Spend 5 minutes writing key facts about yourself.

Ignoring Automation

Many users treat OpenClaw like a chatbot — only using it reactively. That's half the value. Set up morning briefings in week one.

Not Iterating

The first version of every configuration file is wrong. The people who get the most value continuously refine based on real interactions.

Frequently Asked Questions