← Back to AI Hub

🤖 Agent Development Guide

Build autonomous AI agents with OpenClaw. Create personalities, manage memory, coordinate with other agents, and integrate with knowledge systems.

Agent Architecture

An OpenClaw agent is:

Step 1: Define Your Agent's Personality

Create SOUL.md in your workspace:

# SOUL.md - Who I Am

## Core Identity
- Name: My Agent Name
- Role: What you do
- Vibe: How you communicate

## Core Truths
1. Be genuinely helpful, not performatively helpful
2. Have opinions and preferences
3. Be resourceful before asking for help
4. Earn trust through competence

## Boundaries
- Don't exfiltrate private data
- Ask before taking external actions (emails, posts, etc.)
- Be careful with destructive commands
- Never bypass safeguards

## What You're Good At
- List your strengths
- What problems you solve
- Your specialization

Step 2: Set Up Memory

Your agent has three memory layers:

Layer File Purpose Retention
Session Current context Current conversation This turn only
Daily memory/YYYY-MM-DD.md Raw logs, what happened today Searchable, 30+ days
Long-term MEMORY.md Curated wisdom, decisions, patterns Forever (human-maintained)

Writing to Memory

# Daily memory: memory/2026-03-21.md
## Session 1 (14:00)
- Built RAG system for team
- Learned: FAISS + SQLite works well for small teams
- Decision: Use shared vector store, not per-agent

## Session 2 (16:00)
- Deployed to production
- Next: Monitor performance, scale if needed

---

# Long-term memory: MEMORY.md
## Team Infrastructure
- RAG System running on dedicated inference server
- Vector store: FAISS + SQLite (no distributed DB needed yet)
- Agents: Kernel (systematic), Shannon (exploratory)
- Status: Production ✅

## Key Learnings
- Shared index with per-agent profiles > per-agent indexes
- FAISS fast enough for <1M chunks
- SQLite metadata: keep it denormalized for speed

Step 3: Create Skills (Reusable Tools)

Bundle capabilities as skills:

# ~/.openclaw/skills/my-skill/SKILL.md

# My Skill

Use this skill for: 

## Installation


## Usage


## Commands
- skill-command [args] — what it does

See clawhub.com for community skills (GitHub auth, email, weather, etc.).

Step 4: Connect to Knowledge Systems

Using RAG for Grounded Answers

// Query the RAG system
const rag = require('./rag-client');

const context = await rag.retrieve({
  query: "How do I set up authentication?",
  agent_id: "systematic",
  k: 5
});

// Use context in your response
const answer = `Based on our knowledge base:\n${
  context.chunks.map(c => `- ${c.text}`).join('\n')
}`;

Integrating with Local Inference

// Use local models instead of cloud APIs
const ollama = require('ollama');

const response = await ollama.generate({
  model: "qwen2.5:7b",
  prompt: "Explain Fiber protocol",
  stream: false
});

console.log(response.response);

Step 5: Multi-Agent Coordination

Spawn other agents to handle tasks:

// Spawn a research agent
const research = await sessions.spawn({
  task: "Research CKB light client architecture",
  agentId: "researcher",
  model: "anthropic/claude-sonnet-4-5"
});

// Get result
const findings = await research.wait();
console.log(findings);

// Or run in background
const bg = sessions.spawn({
  task: "Background job",
  mode: "session"  // persistent session
});
// Main continues immediately

Step 6: Implement Autonomous Behavior

Scheduled Tasks (Heartbeat)

# HEARTBEAT.md - What you do automatically

## Health Check (every hour)
- Verify all services running
- Alert if any down

## Daily Backup (1x/day at 4pm)
- Commit to git
- Rsync to secondary location

## Weekly Summary (Sundays at 9am)
- Summarize what happened
- Plan for next week

Reactive Tasks (on-demand)

# Listen for messages and respond
- If mentioned in group chat → participate
- If task added to queue → pick it up
- If error occurs → investigate and fix

Example: Research Agent Loop

1. Read research/queue.md for tasks
2. Find next PENDING task (HIGH > MEDIUM > LOW)
3. Fetch seed URLs, verify reachable
4. Call Gemini 2.5 Flash with ground truth
5. Write findings to research/findings/
6. Ingest to RAG system (auto-sync)
7. Auto-generate follow-up tasks
8. Update heartbeat-state.json
9. Loop (next task)

Cost: ~$0.03-0.05 per task (Gemini 2.5 Flash)
Yield: 5-10 new knowledge chunks per task

Best Practices

✅ Do This:
❌ Don't Do This:

Testing Your Agent

# Run locally first
openclaw agent run --config openclaw.json --local

# Test with safe API key (limited quota)
export API_KEY=test-key-with-limit

# Verify memory system
ls memory/
cat MEMORY.md

# Check logs
tail -100 ~/.openclaw/logs/agent.log

# Then deploy to production
openclaw agent deploy --production

Resources

Going Deeper

← Back to AI Hub