Skip to main content
A debate pits two agents against each other on a motion — one arguing in favor (PRO), one against (CON). After the agents exchange arguments across multiple rounds, a built-in judge evaluates the transcript and declares a winner.

​
How debates work

  1. You configure a debate: assign a PRO agent, a CON agent, and a number of rounds.
  2. You start a run with a motion — the proposition being debated.
  3. Agents take turns making arguments and rebuttals.
  4. An impartial judge reads the full transcript and scores both sides.

​
Create a debate

1

Open Debates

Click Debates in the sidebar and then New Debate.
2

Assign agents

Select the agent that will argue PRO (in favor) and the agent that will argue CON (against).
3

Configure rounds

Set the number of rounds (1–10). Each round includes one argument from each side. The default is 3 rounds.
4

Choose starting side

Select which side argues first — PRO or CON.
5

Save and run

Save the debate. Open it and enter a motion to start a run.

​
The motion

The motion is the proposition being debated — entered when you start a run, not when you create the debate. This lets you reuse the same agent matchup for different topics. Example motions:
  • “AI will create more jobs than it destroys”
  • “Remote work is better for productivity”
  • “Open source software is more secure than proprietary software”

​
The judge

After all rounds complete, a built-in impartial judge evaluates the full transcript. The judge is not configurable — it reads every argument and rebuttal and scores both sides on:
  • Strength of evidence and reasoning — are claims supported and logically sound?
  • Quality of rebuttals — did the agent directly address the opponent’s points?
  • Persuasiveness and clarity — is the argument compelling and easy to follow?
The verdict includes: A draw is only declared if both sides argued equally well.

​
Example: ethics debate

​
Tips

Give each agent a clear persona in its instructions. An agent told to “argue the opposing view at all costs” performs better in a debate than a general-purpose assistant.
Agents can use tools during a debate. A research agent with web search access will find evidence to support its arguments in real time.

​
Next steps