Building Multi-Agent Applications with Deep Agents使用 Deep Agents 构建多智能体应用

Using Subagents: Specialized, Isolated Workers使用子智能体：专业、隔离的工作单元

Subagents tackle a fundamental problem in agent engineering: context bloat. This is when an agent's context window becomes close to full as it works on a task.子智能体解决了智能体工程中的一个根本问题：上下文膨胀。当智能体在处理任务时，其上下文窗口接近填满，就会出现这种情况。

Why is this important? There's great work from Chroma on context rot showing that models struggle to complete tasks as their context window gets filled. Our friends at HumanLayer call this high context regime the “dumb zone”. Subagents isolate context from the main agent to help avoid quickly entering the dumb zone.为什么这很重要？Chroma 关于“上下文腐烂 (context rot)”的出色研究表明，当上下文窗口被填满时，模型难以完成任务。我们的朋友 HumanLayer 将这种高上下文状态称为“愚钝区 (dumb zone)”。子智能体将上下文与主智能体隔离开来，有助于避免过快进入“愚钝区”。

When your agent makes dozens of web searches or file reads, the context window fills with intermediate results. Subagents isolate work by running with their own context window. So if the subagent is doing a lot of exploratory work before coming with its final answer, the main agent still only gets the final result, not the 20 tool calls that produced it.当您的智能体进行数十次网络搜索或文件读取时，上下文窗口会被中间结果填满。子智能体通过在各自的上下文窗口中运行来隔离工作。因此，如果子智能体在得出最终答案之前进行了大量的探索性工作，主智能体仍然只会收到最终结果，而不是产生该结果的 20 次工具调用。

Here's a look at the basic subagents architecture:以下是基本子智能体架构的概览：

from deepagents import create_deep_agent
 
research_subagent = {
    "name": "research-agent",
    "description": "Used to research more in depth questions",
    "system_prompt": "You are a great researcher",
    "tools": [internet_search],
    "model": "openai:gpt-4o",  # Optional: override main agent model
}
 
agent = create_deep_agent(
    model="claude-sonnet-4-5-20250929",
    subagents=[research_subagent]
)

research_subagent = {
    "name": "research-agent",
    "description": "Conducts in-depth research using web search and synthesizes findings",
    "system_prompt": """You are a thorough researcher. Your job is to:

    1. Break down the research question into searchable queries
    2. Use internet_search to find relevant information
    3. Synthesize findings into a comprehensive but concise summary
    4. Cite sources when making claims

    Output format:
    - Summary (2-3 paragraphs)
    - Key findings (bullet points)
    - Sources (with URLs)

    Keep your response under 500 words to maintain clean context.""",
    "tools": [internet_search],
}

# ✅ Good: Focused tool set
email_agent = {
    "name": "email-sender",
    # Only email-related
    "tools": [send_email, validate_email],
}
 
# ❌ Bad: Too many tools
email_agent = {
    "name": "email-sender",
    # Unfocused
    "tools": [send_email, web_search, database_query, file_upload],
}

Using Skills: Progressive Disclosure of Capabilities使用技能：渐进式披露能力

Skills provide a different pattern: progressive disclosure. Instead of giving your agent dozens of tools upfront, you define specialized capabilities in SKILL.md files. Your agent sees skill names and descriptions, then reads the full instructions only when needed.技能提供了一种不同的模式：渐进式披露。与其预先给智能体提供数十种工具，不如在 SKILL.md 文件中定义专门的能力。您的智能体可以看到技能名称和描述，然后仅在需要时才读取完整的指令。

Caption: skill descriptions are pre-loaded into the context window. The skill body is only loaded when the agent decides the skill is needed based on the description and previous context.说明：技能描述被预加载到上下文窗口中。技能主体仅在智能体根据描述和先前的上下文决定需要该技能时才会被加载。

.deepagents/skills/
├── deploy/SKILL.md
└── review-pr/SKILL.md

---
name: deploy
description: Deploy to production
version: 1.0.0  # Optional
tags: [deployment, production]  # Optional
---
 
# Deploy to Production
 
When the user asks to deploy, follow these steps:
 
1. Run tests: `npm test`
2. Build the application: `npm run build`
3. Deploy to production: `npm run deploy:prod`
4. Verify deployment: Check the health endpoint
 
Always confirm with the user before deploying to production.

from deepagents import create_deep_agent
from deepagents.backends import FilesystemBackend
 
agent = create_deep_agent(
    model="claude-sonnet-4-5-20250929",
    backend=FilesystemBackend(root_dir="/"),
    skills=[".deepagents/skills"],
)

from deepagents.middleware.filesystem import FileData
 
# default backend is a StateBackend
agent = create_deep_agent(
    model="anthropic:claude-sonnet-4-20250514",
    skills=["/skills/"],
)
 
skill_content = """
---
name: deploy
...
"""
 
# Invoke the agent with the skill and virtual files
result = agent.invoke({
    "messages": [HumanMessage(content="Research the latest Python releases")],
    "files": {
        "/skills/web-research/SKILL.md": FileData(
            content=skill_content.split("\n"),
            created_at="2024-01-01T00:00:00Z",
            modified_at="2024-01-01T00:00:00Z",
        ),
    },
})

Choosing the Right Pattern选择正确的模式

Here's a quick set of questions to guide you:以下是一组快速问题来指导您：

Note, this doesn't have to be an either-or decision.Many systems use both. Skills define procedures; subagents execute complex multi-step work. Your subagents can use skills to effectively manage their context windows!请注意，这不必是二选一的决定。许多系统两者都使用。技能定义程序；子智能体执行复杂的多步骤工作。您的子智能体可以使用技能来有效地管理它们的上下文窗口！

Next Steps后续步骤

To learn more about multi-agent patterns in Deep Agents, check out our:要了解有关 Deep Agents 中多智能体模式的更多信息，请查看我们的：

• Subagents documentation - Detailed API reference and examples子智能体文档 - 详细的 API 参考和示例
• Skills documentation - Detailed API reference and examples技能文档 - 详细的 API 参考和示例
• Multi-agent patterns guide - General guidance on choosing patterns多智能体模式指南 - 关于选择模式的一般指导

The key insight: multi-agent patterns don't have to be complicated. With the right abstractions (middleware for plumbing, tool calling for invocation), they become simple building blocks you can compose into capable, sophisticated systems.关键见解：多智能体模式不必很复杂。有了正确的抽象（用于管道的中间件，用于调用的工具调用），它们就变成了简单的构建块，您可以将其组合成强大、复杂的系统。

Start with subagents for context management, add skills for progressive disclosure, and build from there.从用于上下文管理的子智能体开始，添加用于渐进式披露的技能，并在此基础上进行构建。