Skip to content

gonelake/multi-agent

BranchesTags

Repository files navigation

Multi-Agent 智能写作系统 / Multi-Agent AI Writing System

中文 | English

Demo

中文

一个基于多智能体协作的 AI 内容生成系统，自动完成「热点抓取 → 文章撰写 → 审校修改」全流程，输出适合微信公众号发布的高质量科技文章。

系统架构

ResearcherAgent          WriterAgent              ReviewerAgent
     │                        │                        │
DuckDuckGo 实时搜索    公众号风格写作 Prompt       公众号主编评审标准
     │                        │                        │
     └──────────── Orchestrator 编排层 ────────────────┘
                              │
                    main.py CLI 入口

五层分离架构：

层级 文件 职责
配置层 config.py 所有参数的唯一来源（7 个 Config 类）
框架层 base_agent.py 消息协议、LLMClient、BaseAgent 抽象类
业务层 agents.py Researcher / Writer / Reviewer 三个智能体
编排层 orchestrator.py 工作流调度、审校循环
应用层 main.py CLI 入口、结果输出

核心功能

  • 实时热点抓取：通过 DuckDuckGo 搜索最近 7 天真实新闻，LLM 筛选并推荐最佳选题
  • 公众号风格写作：短段落、口语化、移动端适配，自动生成 A/B/C 三个备选标题
  • 多轮审校循环：ReviewerAgent 按公众号维度评分，WriterAgent 自主评估反馈价值后修改
  • 双格式输出：output.json（完整数据）+ output.md（可直接粘贴的 Markdown 文章）
  • 实验追踪：每次运行自动记录到 experiments.tsv，统计通过率和平均分

快速开始

1. 克隆并安装依赖

git clone <repo-url>
cd multi-agent

pip install -r requirements.txt

如需运行测试，改用 pip install -r requirements-dev.txt

2. 配置 API Key

cp .env.example .env

编辑 .env，填入你的 LLM API Key，并取消注释对应服务商的配置块：

# 必填
LLM_API_KEY=your-api-key-here

# 示例：使用 DeepSeek
LLM_BASE_URL=https://api.deepseek.com/v1
LLM_MODEL=deepseek-chat
LLM_API_STYLE=openai

支持 Kimi、DeepSeek、OpenAI、Claude、Qwen、智谱 GLM、Ollama 等， 完整配置示例见 .env.example 和支持的 LLM API 章节。

3. 验证配置（可选）

# 使用 Mock LLM 快速验证系统运行正常，无需真实 API Key
python main.py --demo

4. 运行

# 默认配置（AI 领域，深度分析，1000 字）
python main.py

# 自定义参数
python main.py --topic 量子计算 --style 科普入门 --words 800
python main.py --topic 机器人 --pass-threshold 80
python main.py --topic 大模型 --max-revisions 3 --output result.json

5. 查看结果

output.md    # Markdown 文章（可直接复制到公众号编辑器）
output.json  # 完整数据（含热点列表、审校历史、评分等）

CLI 参数说明

参数 默认值 说明
--topic AI 搜索领域
--style 深度分析 文章风格
--words 1000 目标字数
--count 5 热点抓取数量
--max-revisions 2 最大审校修改轮次
--pass-threshold 85 审校通过最低分（1-100）
--output output.json JSON 输出文件路径
--description - 本次实验描述（记录到 experiments.tsv）

配置说明

所有参数集中在 config.py，修改后无需传 CLI 参数：

# 修改默认领域和字数
class CliDefaults:
    topic:      str = "AI"
    word_count: int = 1000

# 调整审校评分维度权重
class ReviewConfig:
    dimension_weights = {
        "content_insight": 30,  # 内容洞察
        "readability":     25,  # 可读性
        "title_appeal":    20,  # 标题吸引力
        "structure_flow":  15,  # 结构流畅
        "accuracy":        10,  # 事实准确性
    }

# 支持通过环境变量覆盖的客户端参数
# LLM_TIMEOUT, LLM_MAX_TOKENS, LLM_MAX_RETRIES
# LLM_BASE_URL, LLM_MODEL, LLM_API_STYLE

支持的 LLM API

在 .env 中取消注释对应配置块即可切换，所有 OpenAI 兼容服务商均支持：

# Kimi Code（默认，Anthropic 格式）
LLM_BASE_URL=https://api.kimi.com/coding/v1
LLM_MODEL=kimi-for-coding
LLM_API_STYLE=anthropic

# Kimi（moonshot，OpenAI 兼容）
LLM_BASE_URL=https://api.moonshot.cn/v1
LLM_MODEL=moonshot-v1-8k
LLM_API_STYLE=openai

# DeepSeek
LLM_BASE_URL=https://api.deepseek.com/v1
LLM_MODEL=deepseek-chat
LLM_API_STYLE=openai

# OpenAI
LLM_BASE_URL=https://api.openai.com/v1
LLM_MODEL=gpt-4o
LLM_API_STYLE=openai

# Anthropic Claude
LLM_BASE_URL=https://api.anthropic.com/v1
LLM_MODEL=claude-3-5-sonnet-20241022
LLM_API_STYLE=anthropic

# Qwen（通义千问）
LLM_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1
LLM_MODEL=qwen-plus
LLM_API_STYLE=openai

# 智谱 GLM
LLM_BASE_URL=https://open.bigmodel.cn/api/paas/v4
LLM_MODEL=glm-4-flash
LLM_API_STYLE=openai

# Ollama（本地部署）
LLM_BASE_URL=http://localhost:11434/v1
LLM_MODEL=llama3.2
LLM_API_STYLE=openai
LLM_API_KEY=ollama   # 不需要真实 Key，填任意字符串

完整配置示例见 .env.example

运行测试

# 运行全部测试（不需要真实 API Key）
python -m pytest -v

# 带覆盖率报告
python -m pytest --cov

当前测试状态：156/156 通过，覆盖率 94%

Agent Skill（多 Agent 集成）

本系统提供可跨 agent 使用的 skill，支持 CodeBuddy Code、Claude Code、OpenClaw，安装后 agent 看到写文章、生成新闻等需求时会自动加载并调用本系统。

# 自动检测本机已安装的 agent，逐一安装
bash skills/install.sh

# 或指定安装到特定 agent
bash skills/install.sh --agent codebuddy   # CodeBuddy Code → ~/.codebuddy/skills/
bash skills/install.sh --agent claude      # Claude Code   → ~/.claude/skills/
bash skills/install.sh --agent openclaw   # OpenClaw      → ~/.openclaw/skills/
bash skills/install.sh --agent all        # 强制安装到全部 agent

skill 文档位于 skills/multi-agent-writer/SKILL.md。

项目结构

multi-agent/
├── main.py              # CLI 入口
├── base_agent.py        # 框架层：消息协议 + LLMClient + BaseAgent
├── config.py            # 配置层：7 个 Config 类（带参数校验）
├── agents.py            # 业务层：3 个智能体 + TypedDict 类型定义
├── orchestrator.py      # 编排层：工作流调度
├── search.py            # 搜索层：DuckDuckGo 封装（含客户端时间过滤）
├── experiments.py       # 追踪层：实验日志 TSV 读写
├── .env.example         # API Key 配置模板
├── requirements.txt     # 生产依赖
├── requirements-dev.txt # 开发/测试依赖
├── skills/
│   ├── install.sh                      # skill 一键安装脚本
│   └── multi-agent-writer/
│       └── SKILL.md                    # Agent skill 文档
└── tests/
    ├── conftest.py
    ├── mock_llm.py          # 测试专用 MockLLMClient
    ├── test_base_agent.py
    ├── test_config.py       # 含参数校验测试
    ├── test_agents.py
    ├── test_orchestrator.py
    └── test_experiments.py

工作流示意

1. 热点抓取
   DuckDuckGo 搜索（近 7 天）→ LLM 筛选 5 条热点 → 推荐最佳选题

2. 文章撰写
   选题 → LLM 按公众号风格写作 → 生成正文 + 三个备选标题

3. 审校循环（最多 max_revisions 轮）
   ReviewerAgent 评分 → 通过（≥85）则输出
                       → 未通过则 WriterAgent 评估反馈 → 修改 → 再审校

4. 输出
   output.md（文章）+ output.json（完整数据）

设计亮点

  • 单一真相来源：所有参数集中在 config.py，无散落硬编码
  • 严格 API Key 管理：Key 必须通过 .env 提供，缺失时启动即报错
  • LLM 自修复 JSON：4 级递进解析策略，LLM 输出格式错误时自动重试修复
  • WriterAgent 自主决策：评估审校反馈价值，拒绝无意义反馈，避免无效修改
  • 客户端时间过滤：DuckDuckGo timelimit 参数不可靠，客户端二次过滤确保新闻在时间窗口内
  • 依赖注入：各 Agent 通过构造函数接收 Config，测试时可注入自定义配置

English

A multi-agent AI content generation system that automates the full pipeline of trending topic discovery → article writing → editorial review, producing high-quality tech articles ready for WeChat Official Accounts.

Architecture

ResearcherAgent          WriterAgent              ReviewerAgent
     │                        │                        │
DuckDuckGo Live Search   WeChat-style Prompt      Editorial Review Criteria
     │                        │                        │
     └──────────── Orchestrator (Workflow Layer) ──────┘
                              │
                    main.py CLI Entry Point

5-Layer Separation:

Layer File Responsibility
Config config.py Single source of truth for all parameters (7 Config classes)
Framework base_agent.py Message protocol, LLMClient, BaseAgent abstract class
Business agents.py Researcher / Writer / Reviewer agents
Orchestration orchestrator.py Workflow scheduling, review loop
Application main.py CLI entry point, result output

Features

  • Real-time trending topics: Searches DuckDuckGo for the last 7 days of news; LLM filters and recommends the best topic
  • WeChat-style writing: Short paragraphs, conversational tone, mobile-optimized; auto-generates 3 alternative titles (A/B/C)
  • Multi-round review loop: ReviewerAgent scores by WeChat dimensions; WriterAgent autonomously evaluates feedback value before revising
  • Dual-format output: output.json (full data) + output.md (paste-ready Markdown article)
  • Experiment tracking: Each run is automatically logged to experiments.tsv with pass rate and average score

Quick Start

1. Clone and install dependencies

git clone <repo-url>
cd multi-agent

pip install -r requirements.txt

For running tests, use pip install -r requirements-dev.txt instead.

2. Configure API Key

cp .env.example .env

Edit .env, fill in your API Key and uncomment the block for your provider:

# Required
LLM_API_KEY=your-api-key-here

# Example: DeepSeek
LLM_BASE_URL=https://api.deepseek.com/v1
LLM_MODEL=deepseek-chat
LLM_API_STYLE=openai

Supports Kimi, DeepSeek, OpenAI, Claude, Qwen, Zhipu GLM, Ollama, and more. See .env.example and Supported LLM APIs for all options.

3. Verify setup (optional)

# Run with Mock LLM to verify the system works without a real API Key
python main.py --demo

4. Run

# Default config (AI domain, deep analysis, 1000 words)
python main.py

# Custom parameters
python main.py --topic "quantum computing" --style "popular science" --words 800
python main.py --topic "robotics" --pass-threshold 80
python main.py --topic "LLM" --max-revisions 3 --output result.json

5. View results

output.md    # Markdown article (paste directly into WeChat editor)
output.json  # Full data (trending list, review history, scores, etc.)

CLI Reference

Argument Default Description
--topic AI Search domain
--style 深度分析 Article style
--words 1000 Target word count
--count 5 Number of trending topics to fetch
--max-revisions 2 Maximum review-revision rounds
--pass-threshold 85 Minimum passing score (1-100)
--output output.json JSON output file path
--description - Experiment description (logged to experiments.tsv)

Configuration

All parameters are centralized in config.py:

# Change default domain and word count
class CliDefaults:
    topic:      str = "AI"
    word_count: int = 1000

# Adjust review scoring dimension weights
class ReviewConfig:
    dimension_weights = {
        "content_insight": 30,  # Content insight
        "readability":     25,  # Readability
        "title_appeal":    20,  # Title appeal
        "structure_flow":  15,  # Structure flow
        "accuracy":        10,  # Factual accuracy
    }

# Environment variable overrides
# LLM_TIMEOUT, LLM_MAX_TOKENS, LLM_MAX_RETRIES
# LLM_BASE_URL, LLM_MODEL, LLM_API_STYLE

Supported LLM APIs

Uncomment the corresponding block in .env to switch providers. All OpenAI-compatible services are supported:

# Kimi Code (default, Anthropic format)
LLM_BASE_URL=https://api.kimi.com/coding/v1
LLM_MODEL=kimi-for-coding
LLM_API_STYLE=anthropic

# Kimi (moonshot, OpenAI-compatible)
LLM_BASE_URL=https://api.moonshot.cn/v1
LLM_MODEL=moonshot-v1-8k
LLM_API_STYLE=openai

# DeepSeek
LLM_BASE_URL=https://api.deepseek.com/v1
LLM_MODEL=deepseek-chat
LLM_API_STYLE=openai

# OpenAI
LLM_BASE_URL=https://api.openai.com/v1
LLM_MODEL=gpt-4o
LLM_API_STYLE=openai

# Anthropic Claude
LLM_BASE_URL=https://api.anthropic.com/v1
LLM_MODEL=claude-3-5-sonnet-20241022
LLM_API_STYLE=anthropic

# Qwen
LLM_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1
LLM_MODEL=qwen-plus
LLM_API_STYLE=openai

# Zhipu GLM
LLM_BASE_URL=https://open.bigmodel.cn/api/paas/v4
LLM_MODEL=glm-4-flash
LLM_API_STYLE=openai

# Ollama (local)
LLM_BASE_URL=http://localhost:11434/v1
LLM_MODEL=llama3.2
LLM_API_STYLE=openai
LLM_API_KEY=ollama   # No real key needed, any string works

See .env.example for the full configuration template.

Running Tests

# Run all tests (no real API Key needed)
python -m pytest -v

# With coverage report
python -m pytest --cov

Current test status: 156/156 passing, 94% coverage

Agent Skill (Multi-Agent Integration)

This system ships a skill compatible with CodeBuddy Code, Claude Code, and OpenClaw. Once installed, agents that encounter article-writing or news-generation requests will automatically load and invoke this system.

# Auto-detect installed agents and install to each
bash skills/install.sh

# Or install to a specific agent
bash skills/install.sh --agent codebuddy   # CodeBuddy Code → ~/.codebuddy/skills/
bash skills/install.sh --agent claude      # Claude Code   → ~/.claude/skills/
bash skills/install.sh --agent openclaw   # OpenClaw      → ~/.openclaw/skills/
bash skills/install.sh --agent all        # Force install to all agents

Skill documentation: skills/multi-agent-writer/SKILL.md.

Project Structure

multi-agent/
├── main.py              # CLI entry point
├── base_agent.py        # Framework: message protocol + LLMClient + BaseAgent
├── config.py            # Config: 7 Config classes with validation
├── agents.py            # Business: 3 agents + TypedDict definitions
├── orchestrator.py      # Orchestration: workflow scheduling
├── search.py            # Search: DuckDuckGo wrapper (with client-side time filter)
├── experiments.py       # Tracking: experiment log TSV read/write
├── .env.example         # API Key config template
├── requirements.txt     # Production dependencies
├── requirements-dev.txt # Dev/test dependencies
├── skills/
│   ├── install.sh                      # One-command skill installer
│   └── multi-agent-writer/
│       └── SKILL.md                    # Agent skill documentation
└── tests/
    ├── conftest.py
    ├── mock_llm.py          # MockLLMClient for testing
    ├── test_base_agent.py
    ├── test_config.py       # Includes parameter validation tests
    ├── test_agents.py
    ├── test_orchestrator.py
    └── test_experiments.py

Workflow

1. Trending Topic Discovery
   DuckDuckGo search (last 7 days) → LLM filters 5 topics → recommends best topic

2. Article Writing
   Topic → LLM writes in WeChat style → body + 3 alternative titles

3. Review Loop (up to max_revisions rounds)
   ReviewerAgent scores → pass (≥85) → output
                        → fail → WriterAgent evaluates feedback → revise → re-review

4. Output
   output.md (article) + output.json (full data)

Design Highlights

  • Single source of truth: All parameters in config.py, no scattered hardcoding
  • Strict API Key management: Key must be provided via .env; missing key causes immediate startup error
  • Self-healing JSON parsing: 4-level progressive parsing strategy with automatic retry on LLM format errors
  • WriterAgent autonomy: Evaluates feedback value and rejects low-value suggestions to avoid pointless revisions
  • Client-side time filtering: DuckDuckGo timelimit is unreliable; client-side secondary filter ensures news is within the time window
  • Dependency injection: Each agent receives Config via constructor, enabling custom config injection in tests

License

MIT

About

Agent协同驱动的自主内容生成工具。三个专业AI智能体协同工作，完成从热门话题发现、文章草拟到编辑审阅的端到端流程，产出可直接发布的微信公众号文章。 无需人工干预：一条指令，即可从实时新闻搜索直达精修待发的文章。

Resources

Readme
Activity

Stars

4 stars

Watchers

0 watching

Forks

1 fork
Report repository

Releases 1

多Agent协同，实现AI 自动化内容协作生产引擎 Latest
Mar 27, 2026

Packages

 
 
 

Contributors

Languages