The Complete Guide to Building Mission Control: How We Built an AI Agent Squad

This is the full story of how I built Mission Control. A system where 10 AI agents work together like a real team. If you want to replicate this setup, this guide covers everything. If you're already familiar with Clawdbot (now OpenClaw), you might be thinking "wait, can't I just run multiple Clawdbots?" Yes. That's exactly what this is. This guide shows you how. Part 1: Why I Built This The Problem With AI Assistants I run @SiteGPT, an AI chatbot for customer support. I use AI constantly. But every AI tool I tried had the same problem. No continuity. Every conversation started fresh. Context from yesterday? Gone. That research I asked for last week? Lost in some chat thread I'd never find again. I wanted something different. Agents that remember what they're working on. Multiple agents with different skills working together. A shared workspace where all context lives. The ability to assign tasks and track progress. Basically, I wanted AI to work like a team, not like a search box. The Starting Point: Clawdbot I was already using Clawdbot. It's an open-source AI agent framework that runs as a persistent daemon. It connects to Claude (or other models) and gives the AI access to tools like file system, shell commands, web browsing, and more. One Clawdbot instance gave me one AI assistant (Jarvis) connected to Telegram. Useful, but limited. Then I had a thought. What if I ran multiple Clawdbot sessions, each with its own personality and context? That's when I realized the architecture was already there. I just needed to orchestrate it. Part 2: Understanding Clawdbot Architecture (The Foundation) If you're going to build a multi-agent system, you need to understand how Clawdbot works under the hood. This is the foundation everything else builds on. What Is Clawdbot? Clawdbot (now called OpenClaw) is an AI agent framework with three main jobs: First, it connects AI models to the real world. File access, shell commands, web browsing, APIs. Second, it maintains persistent sessions. Conversation history that survives restarts. Third, it routes messages. Connect the AI to Telegram, Discord, Slack, or any channel. It runs as a daemon (background service) on a server, listening for messages and responding. The Gateway The Gateway is the core process. It runs 24/7 on your server. It manages all active sessions. It handles cron jobs (scheduled tasks). It routes messages between channels and sessions. It provides a WebSocket API for control. Start it with: Configuration lives in a JSON file. You define which AI provider and model to use (Anthropic, OpenAI, etc.), which channels to connect (Telegram, Discord, etc.), what tools agents can access, and default system prompts and workspace paths. Sessions: The Key Concept A session is a persistent conversation with context. Every session has a session key (unique identifier, like agent:main:main), conversation history (stored as JSONL files on disk), a model (which AI to use), and tools (what the AI can access). Here's the important part. Sessions are independent. Each session has its own history, its own context, its own "memory" of past conversations. When you run multiple agents, you're really running multiple sessions. Each with their own identity. How Sessions Work Sessions can be main sessions (long-running, interactive, like chatting with Jarvis) or isolated sessions (one-shot, for cron jobs, wake up, do task, done). Cron Jobs: Scheduled Agent Wakeups Clawdbot has a built-in cron system. You can schedule tasks: When a cron fires, the Gateway creates or wakes a session, sends the message to the AI, the AI responds (can use tools, send messages, etc.), and the session can persist or terminate. This is how agents "wake up" periodically without being always-on. The Workspace Every Clawdbot instance has a workspace. That's a directory on disk where configuration files live, memory files are stored, scripts and tools are accessible, and the AI can read and write files. The workspace is how agents persist information between sessions. They write to files. Those files survive restarts. Part 3: From One Clawdbot to Ten Agents Now you understand the foundation. Here's how I built a team. The Insight Clawdbot sessions are independent. Each can have its own personality (via SOUL.md), its own memory files, its own cron schedule, its own tools and access. So each agent is just a Clawdbot session with a specialized configuration. Jarvis isn't special. He's a session with session key agent:main:main, a SOUL.md that says "You are Jarvis, the squad lead...", access to all tools, and a connection to my Telegram. Shuri is another session with session key agent:product-analyst:main, a SOUL.md that says "You are Shuri, the product analyst...", the same tools (file access, shell, browser), and her own heartbeat cron. Ten agents equals ten sessions. Each waking up on their own schedule. Each with their own context. Session Keys: Agent Identity Each agent has a unique ses