Most multi-agent systems I've seen on Twitter share one fatal trait: they try to be smart. The agent reasons about what to do, then does it. The result is a fragile chain — one wrong reasoning step and the entire pipeline derails.
OpenClaw goes the other way. Every agent is a specialist with a small, well-defined job. The Coder writes code. The Git agent commits and pushes. The Deploy agent runs the Vercel CLI. None of them know what the others do. They share state through Firebase, not through prompts.
The three agencies
I split the work along the boundaries where a real org would: Dev (Shuri as CEO), LinkedIn (Natasha), and Creative (Maya). Each agency has its own CEO and its own sub-agents. Above them sits Jarvis, the COO — the single point of contact.
When you give Jarvis a task, Jarvis decides which agency it belongs to. Inside that agency, the CEO routes to sub-agents. The CEO knows the agency's playbook; the sub-agents know nothing outside their narrow role.
This sounds bureaucratic. It is bureaucratic. That's the point — bureaucracy is what keeps a system reliable when no one in it can think the whole thing through.
Specialisation beats reasoning
Early on I tried building one big "product engineer" agent that handled everything from code to deploy. It worked in demos and failed in production within hours. The problem: when you let an agent reason across domains, every domain becomes a place it can hallucinate.
Narrowing the scope to "only write valid TypeScript for this exact file" or "only commit with this exact message format" cut hallucinations sharply. The agents are dumber. The system is smarter.
Shared state, not shared prompts
The other unlock was state. Most multi-agent demos pass context via prompts — Agent A's output becomes Agent B's input. That's how you get drift: by step 5 the context has accumulated five rounds of paraphrasing.
OpenClaw agents read and write to Firebase. They share state, not narrative. Each agent looks at the same source of truth and contributes its slice. The COO arbitrates conflicts.
It's not glamorous. It's not even particularly clever. But it ran a launch campaign that hit 841 impressions without me touching the keyboard, so I'll take it.
— Saurabh