多代理辯論
PR Ready結構化辯論,包含正反方代理、主持人綜合,以及人工裁決。
7 個節點 · 7 條連接pr ready
autogenmulti-agentdebatereasoning
視覺化
提出辯題human
使用者提供辯論主題與任何限制條件。
↓sequential→ 主持人 — 開場
主持人 — 開場agent
闡明問題、定義範疇並設定辯論規則。
↓parallel→ 正方代理
↓parallel→ 反方代理
正方代理agent
以證據與推理為命題進行正方論述。
↓sequential→ 反駁回合
反方代理agent
以反駁證據對命題進行反方論述。
↓sequential→ 反駁回合
反駁回合agent
雙方代理各自回應對方論點(共 2 輪)。
↓sequential→ 主持人 — 綜合
主持人 — 綜合agent
總結雙方論點,點出最強論據與論述缺口。
↓sequential→ 人工裁判
人工裁判human
審閱辯論記錄,宣布勝方或平局。
ex-autogen-debate.osop.yaml
# AutoGen Multi-Agent Debate — OSOP Portable Workflow
#
# A structured debate between AI agents: a moderator poses a question,
# pro and con agents argue in parallel, the moderator synthesizes both
# sides, and a human judge delivers the final verdict.
#
# Run with AutoGen or validate: osop validate autogen-debate.osop.yaml
osop_version: "1.0"
id: "autogen-debate"
name:"多代理辯論"
description:"結構化辯論,包含正反方代理、主持人綜合,以及人工裁決。"
version: "1.0.0"
tags: [autogen, multi-agent, debate, reasoning]
nodes:
- id: "pose_question"
type: "human"
name: "提出辯題"
description: "使用者提供辯論主題與任何限制條件。"
- id: "moderator_open"
type: "agent"
subtype: "llm"
name: "主持人 — 開場"
description: "闡明問題、定義範疇並設定辯論規則。"
config:
role: "Neutral Moderator"
model: "gpt-4o"
- id: "pro_agent"
type: "agent"
subtype: "llm"
name: "正方代理"
description: "以證據與推理為命題進行正方論述。"
config:
role: "Proponent"
tools: [web_search, cite_source]
- id: "con_agent"
type: "agent"
subtype: "llm"
name: "反方代理"
description: "以反駁證據對命題進行反方論述。"
config:
role: "Opponent"
tools: [web_search, cite_source]
- id: "rebuttal_round"
type: "agent"
subtype: "llm"
name: "反駁回合"
description: "雙方代理各自回應對方論點(共 2 輪)。"
config:
max_iterations: 2
- id: "moderator_synthesize"
type: "agent"
subtype: "llm"
name: "主持人 — 綜合"
description: "總結雙方論點,點出最強論據與論述缺口。"
- id: "human_judge"
type: "human"
subtype: "review"
name: "人工裁判"
description: "審閱辯論記錄,宣布勝方或平局。"
edges:
- from: "pose_question"
to: "moderator_open"
mode: "sequential"
- from: "moderator_open"
to: "pro_agent"
mode: "parallel"
- from: "moderator_open"
to: "con_agent"
mode: "parallel"
- from: "pro_agent"
to: "rebuttal_round"
mode: "sequential"
- from: "con_agent"
to: "rebuttal_round"
mode: "sequential"
- from: "rebuttal_round"
to: "moderator_synthesize"
mode: "sequential"
- from: "moderator_synthesize"
to: "human_judge"
mode: "sequential"