56 lines
2.2 KiB
Markdown
56 lines
2.2 KiB
Markdown
---
|
||
title: "Agent Specialization"
|
||
type: concept
|
||
tags: [OpenClaw, Agent, Architecture, Team]
|
||
sources: [multi-agent-team]
|
||
last_updated: 2026-04-23
|
||
---
|
||
|
||
## Definition
|
||
|
||
**Agent Specialization** 是多 Agent 系统中,每个 Agent 拥有独立角色、专长领域和针对性优化的模型,通过协作完成复杂任务的架构设计原则。
|
||
|
||
## 核心洞察
|
||
|
||
> "One agent can't do everything well: A single agent's context window fills up fast when juggling strategy, code, marketing research, and business analysis"
|
||
|
||
核心问题:
|
||
- **上下文窗口瓶颈**:单个 Agent 同时处理多领域任务时上下文快速填满
|
||
- **泛化输出的局限**:通用提示产生通用输出,专业任务需要专业 Agent
|
||
- **效率不均衡**:不同任务需要的模型能力不同,混用造成资源浪费
|
||
|
||
## 专业化 Agent 示例
|
||
|
||
| Agent | 角色 | 模型选择 | 核心职责 |
|
||
|-------|------|----------|----------|
|
||
| Milo | 策略领导 | Claude Opus | 战略规划、OKR 追踪、Agent 协调 |
|
||
| Josh | 商业分析 | Claude Sonnet | 定价策略、KPI 追踪、竞品分析 |
|
||
| Marketing | 营销研究 | Gemini | 内容创意、SEO 研究、社媒监控 |
|
||
| Dev | 开发工程 | Claude Opus/Codex | 代码审查、架构决策、文档撰写 |
|
||
|
||
## 模型匹配原则
|
||
|
||
> "Right model for the right job: Don't use an expensive reasoning model for keyword monitoring"
|
||
|
||
- **复杂推理任务**:Claude Opus(深度思考、架构设计)
|
||
- **分析性任务**:Claude Sonnet(快速分析、数据处理)
|
||
- **研究性任务**:Gemini(长上下文、网页研究)
|
||
- **执行性任务**:Codex(代码生成、工具调用)
|
||
|
||
## 起步策略
|
||
|
||
> "Start with 2, not 4: Begin with a lead + one specialist, then add agents as you identify bottlenecks"
|
||
|
||
从最小可行团队开始:
|
||
1. **1 个领导 Agent**:总协调、决策、汇总
|
||
2. **1 个专家 Agent**:根据当前最大瓶颈选择
|
||
3. **逐步扩展**:瓶颈出现时添加新 Agent
|
||
|
||
## Related Concepts
|
||
|
||
- [[Agent Personality]] — Agent 个性化与专业化相辅相成
|
||
- [[Multi-Agent Coordination]] — 多 Agent 协调机制
|
||
- [[Parallel Agent Execution]] — 并行执行提升效率
|
||
- [[Chained Agents]] — 链式 Agent(另一种协作模式)
|
||
|