How I Built My 10-Agent OpenClaw Team

What It Covers

The host details building a 10-agent OpenClaw team as a nontechnical user, covering agent architecture, practical use cases, setup process, and real-world lessons. He explains which agents deliver value (research, task management) versus which underperform (builder bot), and emphasizes using Claude as a build partner throughout the entire process.

Key Questions Answered

• •Agent Architecture Fundamentals: Each OpenClaw agent loads markdown files at startup defining its identity (name, emoji, description), soul.md (personality and behavior), agents.md (operating protocols), user.md (everything about you), tools.md (system access), and memory.md (long-term learnings). Heartbeat.md fires every 30 minutes for autopilot tasks, while cron jobs schedule specific times like 8AM status updates or 5PM check-ins for end-of-day additions.
• •Research Agents for Continuous Work: Two dedicated research agents run 24/7 surfacing studies and surveys for AIDB Intelligence products, actively proposing changes to maturity maps and opportunity radars based on findings. This requires quality calibration to distinguish good versus poor resources and improve proposal writing, but heartbeats can be flaky and occasionally drop off requiring resets. The persistent sweep of information delivers clear value despite technical hiccups.
• •Project Manager Evolution Strategy: Four project manager agents currently function as glorified to-do list managers, receiving brain dumps about challenges and decisions, then sending reminders (including skull emoji piles every 30 minutes as snooze buttons). Phase two will expand their remit to interact with other systems via skills like Slack access and communicate with other team members' agents, transforming them from personal assistants into true coordinators.
• •Build Partner Methodology: Using Claude Projects with dozens of messages and context files eliminates the need for YouTube tutorials or web guides. Tell Claude your skill level (even "incompetent neophyte") and it provides step-by-step guidance with infinite patience until context window limits appear. Every prompt, problem, and simple command goes into the chat, enabling nontechnical users to build complete agent teams without prior coding experience.
• •Mac Mini Server Setup: A dedicated Mac Mini provides a fresh environment with controlled system access, always-on operation, and remote access from anywhere. Install Homebrew package manager, Node.js, Cloud Code, disable sleep mode, and set up Tailscale for private network access. This approach costs money but eliminates security concerns about bleeding into existing systems, though any laptop works fine for basic OpenClaw deployment.