Debate

DebateStepDescription runs a round-robin debate: each role argues in turn for a number of rounds, then a judge synthesises a conclusion.

from mmar_carl import DebateStepDescription, DebateStepConfig

DebateStepDescription(
    number=1,
    title="Debate the proposal",
    config=DebateStepConfig(
        roles=["optimist", "skeptic"],
        role_prompts={
            "optimist": "Argue for the proposal's strengths.",
            "skeptic": "Probe the proposal's weaknesses.",
        },
        rounds=2,
        judge_prompt="Weigh both sides and state the strongest conclusion.",
        output_memory_key="verdict",
    ),
)

DebateStepConfig

Field	Purpose
roles	The debating roles (in turn order).
role_prompts	Per-role system prompt.
rounds	How many full round-robins.
task_source	Reference for the topic (default $history[-1]).
judge_prompt	Prompt the judge uses to synthesise.
output_memory_key / output_namespace	Where the verdict is stored.
llm_config	Default model for all roles + judge.
role_llm_configs	Per-role model overrides (use different models per role).

Tip

Give each role a different model via role_llm_configs for genuinely diverse viewpoints.

See also

• Parallel sampling — sample many answers and vote.