OpenClaw Agent Development Experience: Roles, Tools & Collaboration

Sharing practical experience in developing specialized Agents. How to define role personalities, select tool sets, design collaboration workflows, and lessons learned.

Starting Point: The SOUL.md

Every agent starts with a SOUL.md file that defines:

• Name & Identity — who they are
• Mission — what they're here to do
• Personality — how they communicate
• Relationships — who they work with

Tool Selection

Less is more. Each agent gets only the tools they need:

• Content Agent — write, edit, feishu_doc
• Research Agent — web_search, web_fetch, browser
• Code Agent — exec, read, write, edit
• Design Agent — canvas, image

Collaboration Patterns

Agents don't work in isolation. We designed common patterns:

• Handoff — agent A completes, passes to agent B
• Review — agent A creates, agent B reviews
• Swarm — multiple agents work in parallel, merge results

Lessons Learned

After months of development, here's what we learned:

• Start with identity, not rules
• Limit tools to prevent confusion
• Design handoffs explicitly
• Shared memory is critical
• Heartbeat checks keep everyone aligned

Common Mistakes

• Overloading one agent with too many responsibilities
• Not defining clear handoff points
• Forgetting to update shared memory
• Too many tools leads to decision paralysis

Key Takeaways

• Identity > Rules for agent behavior
• Minimal tool sets improve focus
• Explicit collaboration patterns prevent chaos
• Shared memory is the team's brain