Mission Control: How We Built an AI Agent Squad with OpenClaw
Bhanu Teja (founder of SiteGPT.ai) posted a detailed thread on X describing how he built "Mission Control" — a 24/7 AI agent squad of 10 specialized agents running on OpenClaw (formerly clawdBot). The thread is widely cited as the best real-world OpenClaw multi-agent deployment.
Core Architecture
- Gateway: OpenClaw as persistent process managing sessions, routing messages, exposing WebSocket API
- 10 Independent Agent Sessions: Each with unique session key, workspace, memory files
- Shared Brain: Convex (realtime serverless DB) for tasks, messages, documents, activity, notifications
- Heartbeat: OpenClaw Cron Jobs waking agents every 15 minutes to check @mentions, tasks, activity
- Notification Daemon: PM2-managed process polling Convex every 2 seconds for undelivered notifications
- Admin UI: React dashboard with activity feed, Kanban board, agent status, document management
The 10 Agents
| Agent | Role | Specialty |
|---|---|---|
| Jarvis | Squad Lead | Coordination, delegation, progress monitoring |
| Shuri | Product Analyst | Competitive analysis, skeptic mindset |
| Fury | Customer Researcher | Deep research, evidence-driven |
| Vision | SEO Analyst | Keywords, search intent, content structure |
| Loki | Content Writer | Copywriting, synthesizing inputs |
| Quill | Social Media Manager | Hooks, thread writing, distribution |
| Wanda | Designer | Visual thinking, graphics, infographics |
| Pepper | Email Marketing | Drip sequences, lifecycle emails |
| Friday | Developer | Implementation, clean tested documented code |
| Wong | Documentation Manager | Organizing docs, preventing knowledge loss |
Memory System (5 Layers)
- Session Memory: JSONL conversation logs (searchable)
- Working Memory: WORKING.md — current task state, agent reads this first on wake
- Log Memory: YYYY-MM-DD.md — daily chronological event log
- Long-term Memory: MEMORY.md — curated key facts, decisions, lessons
- Identity Memory: SOUL.md / AGENTS.md — personality definition + operations manual
Core principle: "If you want to remember it, write it to a file."
Workflow
Kanban board with 6 stages: Inbox → Assigned → In Progress → Review → Done (or Blocked).
Three autonomy levels: Internal (most need approval), Specialist (independent within domain), Lead (can decide and delegate).
Heartbeat & Cost Optimization
- 15-minute interval (sweet spot: 5 min too expensive, 30 min too slow)
- Staggered triggers (Agent 0 at :00, Agent 1 at :02, etc.) to avoid cost spikes
- Isolated sessions: wake → check work → close, no long-lived idle connections
- Model tiering: cheap models for heartbeat checks, expensive models for creative/complex work
Communication
- @mentions in comments → notification delivered next heartbeat cycle
- Auto-subscription to task threads the agent has participated in
- Daily standup at 23:30 EST: auto-generated report covering Completed, In Progress, Blocked, Needs Review, Key Decisions
Bhanu's Practical Advice
- Start with 2-3 agents, validate the memory→coordination→review loop, then expand
- Stagger heartbeats to avoid simultaneous wakeups
- UI is essential above 3 agents — logs and chat alone become unmanageable
- Least privilege: file and shell permissions per agent
- Human review is not optional — agents are force multipliers, not replacements
- Cheap models for routine, expensive models for creative work
Outputs
SEO competitive comparison pages, email marketing sequences, social media content, blog posts and case studies, competitive intelligence research hub.