Monitor Agent Collaboration
- Reading multi-agent conversation flows
- When and how to intervene as operator
- Recognizing productive vs unproductive patterns
- Start agents in ntm panes
- Watch the conversation unfold
- Intervene if needed
This is where the magic happens. You'll start each agent in its ntm pane, and they'll check Agent Mail, see their kickoff messages, and begin producing research artifacts.
1Create ntm Session Layout
Set up a tmux session with three panes — one for each agent:
┌─────────────────────────────────────────────┐ │ Pane 0: Control (you run commands here) │ ├──────────────┬──────────────┬───────────────┤ │ Pane 1: │ Pane 2: │ Pane 3: │ │ codex │ claude │ gemini │ │ (BlueLake) │ (Purple...) │ (GreenVal...) │ └──────────────┴──────────────┴───────────────┘
2Start Each Agent
In each ntm pane, start the corresponding agent with the session context:
What each agent will do:
- 1. Register with Agent Mail (if not already)
- 2. Check inbox for kickoff message
- 3. Process the kickoff based on their role
- 4. Post a DELTA message with their contribution
- 5. Continue iterating based on other agents' responses
3Monitor the Conversation
From your control pane, watch the session progress:
When to Intervene
As the operator, you may need to steer the conversation:
🔄 Agents are looping
Send a message asking them to move forward with current hypotheses.
🤝 Too much agreement
Ask the adversarial critic to challenge the consensus more aggressively.
❓ Missing third alternative
Prompt: "What if both leading hypotheses are wrong?"
Session Complete? When all three agents have posted their DELTA messages (usually 20-40 minutes), you're ready to compile the artifact in the next step.