wecomcli-workflow-multi-agent-dev

WeCom Multi-Programming-Agent Collaboration

Launch multiple sub-agents to work in parallel via Agent Tool, track tasks with WeCom smart sheets, and notify results via messages.

Initial Configuration (Automatic Initialization)

Before performing any operation, first check whether a multi-agent collaboration configuration already exists in memory. If not, automatically create it according to the following process:

Step 1: Create a smart sheet

wecom-cli doc create_doc '{"doc_type": 10, "doc_name": "多代理开发任务"}'

Obtain from the returned result docid。

Step 2: Create sub-sheets and fields

# 创建子表1: 开发任务
wecom-cli doc smartsheet_add_sheet '{"docid": "<DOCID>", "sheet_name": "开发任务"}'

# 创建子表2: 竞赛记录
wecom-cli doc smartsheet_add_sheet '{"docid": "<DOCID>", "sheet_name": "竞赛记录"}'

Development task fields:

wecom-cli doc smartsheet_add_fields '{"docid": "<DOCID>", "sheet_id": "<SHEET_ID>", "fields": [
  {"field_title": "任务描述", "field_type": "FIELD_TYPE_TEXT"},
  {"field_title": "工作模式", "field_type": "FIELD_TYPE_SINGLE_SELECT"},
  {"field_title": "分配代理", "field_type": "FIELD_TYPE_TEXT"},
  {"field_title": "状态", "field_type": "FIELD_TYPE_SINGLE_SELECT"},
  {"field_title": "结果摘要", "field_type": "FIELD_TYPE_TEXT"},
  {"field_title": "耗时", "field_type": "FIELD_TYPE_TEXT"},
  {"field_title": "创建时间", "field_type": "FIELD_TYPE_TEXT"}
]}'

Competition record fields:

wecom-cli doc smartsheet_add_fields '{"docid": "<DOCID>", "sheet_id": "<SHEET_ID>", "fields": [
  {"field_title": "任务描述", "field_type": "FIELD_TYPE_TEXT"},
  {"field_title": "代理1方案", "field_type": "FIELD_TYPE_TEXT"},
  {"field_title": "代理1评分", "field_type": "FIELD_TYPE_TEXT"},
  {"field_title": "代理2方案", "field_type": "FIELD_TYPE_TEXT"},
  {"field_title": "代理2评分", "field_type": "FIELD_TYPE_TEXT"},
  {"field_title": "推荐方案", "field_type": "FIELD_TYPE_SINGLE_SELECT"},
  {"field_title": "对比分析", "field_type": "FIELD_TYPE_TEXT"},
  {"field_title": "创建时间", "field_type": "FIELD_TYPE_TEXT"}
]}'

Step 3: Save the configuration to memory

Save the following information to the memory file multi-agent-dev-config.md：

---
name: multi-agent-dev-config
description: 多代理协同智能表格配置信息
type: project
---

## 多代理协同表配置

- docid: <DOCID>
- url: <创建时返回的URL>
- 表格名称: 多代理开发任务
- 开发任务 sheet_id: <SHEET_ID>
- 竞赛记录 sheet_id: <SHEET_ID>

---

Operation Process

In the following commands, <DOCID> and each <SHEET_ID> are read from multi-agent-dev-config in memory.

Mode 1: Competition Mode

The same task is assigned to 2-3 agents to complete independently, and the AI compares and selects the best solution.

Applicable scenarios: Need the optimal solution, architecture selection, algorithm implementation

Interaction flow:

1. User: "Use competition mode to help me implement a user login module, with two agents each writing one version"
2. AI records the task in the smart sheet
3. Use Agent Tool to launch 2 sub-agents (subagent_type: "general-purpose"）
4. The two agents each complete the implementation
5. AI compares the two solutions, scores them, and selects the best
6. Send the result to the message

Agent Tool call example:

使用 Agent Tool 并行启动 2 个 general-purpose 子代理：

代理1 Prompt：实现一个用户登录模块。要求：
- 使用 JWT + Refresh Token 方案
- 包含密码强度校验
- 包含输入验证和错误处理
- 代码规范、有注释

代理2 Prompt：实现一个用户登录模块。要求：
- 使用 Session-based 方案
- 包含密码强度校验
- 包含输入验证和错误处理
- 代码规范、有注释

Record result:

# 记录各代理任务
wecom-cli doc smartsheet_add_records '{"docid": "<DOCID>", "sheet_id": "<开发任务SHEET_ID>", "records": [
  {"values": {"任务描述": [{"type":"text","text":"用户登录模块"}], "工作模式": [{"text":"竞赛"}], "分配代理": [{"type":"text","text":"Agent1"}], "状态": [{"text":"已完成"}], "结果摘要": [{"type":"text","text":"JWT方案，完整性好"}], "耗时": [{"type":"text","text":"45s"}], "创建时间": [{"type":"text","text":"2026-04-02"}]}},
  {"values": {"任务描述": [{"type":"text","text":"用户登录模块"}], "工作模式": [{"text":"竞赛"}], "分配代理": [{"type":"text","text":"Agent2"}], "状态": [{"text":"已完成"}], "结果摘要": [{"type":"text","text":"Session方案，简洁"}], "耗时": [{"type":"text","text":"38s"}], "创建时间": [{"type":"text","text":"2026-04-02"}]}}
]}'

# 记录竞赛对比
wecom-cli doc smartsheet_add_records '{"docid": "<DOCID>", "sheet_id": "<竞赛记录SHEET_ID>", "records": [{"values": {
  "任务描述": [{"type":"text","text":"用户登录模块"}],
  "代理1方案": [{"type":"text","text":"JWT + Refresh Token，完整性好"}],
  "代理1评分": [{"type":"text","text":"8.5"}],
  "代理2方案": [{"type":"text","text":"Session-based，简洁清晰"}],
  "代理2评分": [{"type":"text","text":"7.0"}],
  "推荐方案": [{"text":"代理1"}],
  "对比分析": [{"type":"text","text":"JWT方案更符合现代最佳实践，Session方案更简洁但缺少密码强度校验"}],
  "创建时间": [{"type":"text","text":"2026-04-02"}]
}}]}'

Mode 2: Division-of-Labor Mode

Different tasks are assigned to different agents for parallel execution.

Applicable scenarios: Multiple independent tasks, batch bug fixes, parallel feature development

Interaction flow:

1. User: "I have 3 tasks to do in parallel: fix bug #123, write the login page, optimize the SQL query"
2. AI launches one sub-agent for each task
3. All agents execute in parallel
4. Summarize the results

Agent Tool call:

并行启动 3 个 general-purpose 子代理：

代理1：修复 bug #123 — <具体 bug 描述>
代理2：写登录页面 — <具体需求>
代理3：优化 SQL 查询 — <具体优化目标>

Record result:

wecom-cli doc smartsheet_add_records '{"docid": "<DOCID>", "sheet_id": "<开发任务SHEET_ID>", "records": [
  {"values": {"任务描述": [{"type":"text","text":"修bug #123"}], "工作模式": [{"text":"分工"}], "分配代理": [{"type":"text","text":"Agent1"}], "状态": [{"text":"已完成"}], "耗时": [{"type":"text","text":"45s"}], "创建时间": [{"type":"text","text":"2026-04-02"}]}},
  {"values": {"任务描述": [{"type":"text","text":"写登录页面"}], "工作模式": [{"text":"分工"}], "分配代理": [{"type":"text","text":"Agent2"}], "状态": [{"text":"已完成"}], "耗时": [{"type":"text","text":"120s"}], "创建时间": [{"type":"text","text":"2026-04-02"}]}},
  {"values": {"任务描述": [{"type":"text","text":"优化SQL查询"}], "工作模式": [{"text":"分工"}], "分配代理": [{"type":"text","text":"Agent3"}], "状态": [{"text":"已完成"}], "耗时": [{"type":"text","text":"60s"}], "创建时间": [{"type":"text","text":"2026-04-02"}]}}
]}'

Send summary notification:

wecom-cli msg send_message '{"chat_type": 1, "chatid": "<USERID>", "msgtype": "text", "text": {"content": "分工任务全部完成！\n\n| 任务 | 代理 | 状态 | 耗时 |\n|------|------|------|------|\n| 修bug #123 | Agent1 | 完成 | 45s |\n| 写登录页面 | Agent2 | 完成 | 120s |\n| 优化SQL查询 | Agent3 | 完成 | 60s |\n\n总耗时 120s（串行需 225s，加速 1.9x）"}}'

Mode 3: Pipeline Mode

Code → review → test, completed in sequence by different agents collaborating.

Applicable scenarios: Complex feature development, tasks requiring quality assurance

Interaction flow:

1. User: "Use pipeline mode to help me refactor this module"
2. Agent 1: Implement/refactor the code
3. Agent 2: Code review, propose improvement suggestions
4. Agent 3: Add tests (optional)
5. Summarize the results

Agent Tool call:

按顺序启动 3 个子代理（注意：流水线是串行的，不是并行）：

阶段1 代理：重构 <模块名> 模块的代码。要求：<具体要求>

阶段2 代理（接收阶段1结果）：对以下代码进行审查。要求：
- 检查代码质量、潜在 bug、性能问题
- 提出 3-5 个改进建议
- 评估代码可维护性

阶段3 代理（接收阶段1+2结果）：为以下代码编写完整的单元测试。要求：
- 覆盖所有主要功能路径
- 包含边界情况测试

---

Comparison of the Three Modes

Mode	Applicable scenarios	Number of agents	Execution method	Speedup ratio
Competition	Need the optimal solution	2-3	Same task, completed separately, AI comparison	1x (take the slowest)
Division of labor	Multiple independent tasks	By number of tasks	Different tasks, completed in parallel	Nx
Pipeline	Complex features, require quality assurance	2-3	Code → review → test, sequential collaboration	~0.7x (with review overhead)

---

Field type write format (general rules)

For the complete field type format description, see crm-schema

Field type	Write format	Example
TEXT	[{"type":"text","text":"值"}]	Task description, agent name, duration, etc.
SINGLE_SELECT	[{"text":"选项文本"}]	Work mode, status, recommended solution, etc.

---

Practical notes

Common pitfalls

1. Not passing key_type when updating — must include "key_type": "CELL_VALUE_KEY_TYPE_FIELD_TITLE"
2. When an API call fails — check the error code: 301085 means the docid is invalid (use url instead); 2022004 means the field was not found (call add_fields first); 2022003 means the record does not exist (call add_records first). When encountering an error, first use smartsheet_get_sheet '{"url": "<URL>"}' to verify the table structure.
3. Agent Tool isolation — the two agents in competition mode need to work independently in isolation: "worktree" to avoid interfering with each other
4. Pipeline serial dependency — in pipeline mode, the prompts for stages 2 and 3 need to include the output of stage 1
5. Task granularity — each task in division-of-labor mode should be sufficiently independent to avoid cross-dependencies
6. Agent count limit — it is recommended to start at most 3 agents at the same time to avoid resource contention
7. options do not take effect when creating SINGLE_SELECT — smartsheet_add_fields passed in property.options will not pre-create options; options are automatically created via [{"text": "选项"}] when a record is first written. When creating the field, the property parameter can be omitted.

Data linkage

• This Skill runs independently and does not depend on other workflow Skills
• Use the underlying wecom-cli todo to create tracking to-dos (optional)
• Use the underlying wecom-cli msg to send notifications