OpenClaw: How We Got Three AI Agents to Debate Each Other
We built a Multi-Agent Debate system with OpenClaw: Three AI agents discuss in separate terminal sessions until they reach consensus. Here's what happened.
In the first article about OpenClaw, we covered the basic idea: an AI assistant that runs on your own hardware and is controllable via WhatsApp. Since then, we’ve been experimenting further — and built something that shows where AI automation is actually heading.
We got three AI agents to debate each other in separate terminal sessions — until they agreed on a solution.
The Problem: A Single Agent Isn’t Enough
Anyone who works with ChatGPT or Claude knows this: the first answer sounds good. But when you probe deeper, gaps appear. The model confirms itself, overlooks risks, stays trapped in its own logic.
Current research (ICML 2024, arXiv 2025) shows: when multiple agents with different roles work on the same problem and critique each other, hallucinations decrease measurably. Results become more robust.
The concept is called Multi-Agent Debate (MAD). And it’s not just theory — we implemented it on our own infrastructure.
What We Built: The MAD Controller
Specifically, we implemented a MAD orchestrator based on OpenClaw and tmux. tmux is a terminal multiplexer — software that allows multiple terminal sessions to run in parallel without requiring an active connection.
Our setup: A tmux session with three panes, each running a specialized agent:
Pane 0 — The Generator: Creates concrete, practical solution proposals. Takes previous criticism into account and improves suggestions round by round.
Pane 1 — The Critic: Reads the Generator’s proposal, analyzes it for weaknesses, missing aspects, and practical feasibility. Constructive but sharp.
Pane 2 — The Judge: Monitors the exchange between Generator and Critic. Evaluates whether arguments are converging. And decides when stable consensus is reached.
The agents communicate via tmux capture-pane — the orchestrator reads each pane’s output, passes it to the next agent, and writes status messages into the respective sessions. Maximum five rounds, early termination on consensus.
Why tmux Is the Key
Most AI tools execute commands once: command in, result out, done. Context is lost.
tmux solves this. Sessions run persistently in the background. The agent can send a command, read the output, reason about it, and issue a follow-up command in the same session. This simulates how a human works at the terminal — just without pausing between steps.
For the MAD controller, this means: Three agents work simultaneously in their own environments, see each other’s results, and react accordingly. All local, all on your own hardware, no cloud dependency.
The Result: Consensus After Three Rounds
Our test topic was: “How can a small business meaningfully use AI in day-to-day operations?”
What happened:
- Round 1: The Generator delivered a solid but generic proposal. The Critic flagged missing cost estimates and unrealistic expectations about SMB resources.
- Round 2: The Generator revised — more specific, with budget ranges and prioritization. The Critic identified a gap around data privacy.
- Round 3: The Generator added GDPR-compliant alternatives. The Judge recognized consensus and delivered a synthesis.
The result was significantly more nuanced than any single ChatGPT response. Not because the model was better — but because the architecture forced it to address weaknesses instead of ignoring them.
Why This Matters for Businesses
This experiment demonstrates a principle that goes beyond debating:
Agents that check each other deliver better results than agents working alone. This applies to text generation just as much as to code reviews, compliance checks, or strategic analysis.
And with tmux as the persistence layer and OpenClaw as the orchestration layer, these multi-agent systems can run on your own infrastructure — without sending data to third parties, without vendor lock-in.
The question is no longer “Can AI do this?” but: How do you orchestrate it properly?
Want to know how AI agents can be deployed in your organization? We advise on architecture, security, and integration — pragmatically, without hype.
Sources: openclaw.ai | Research on Multi-Agent Debate: ICML 2024, arXiv 2025