Subagents

Learn how to use subagents to delegate work and keep context clean

Deep Agents can create subagents to delegate work. You can specify custom subagents in the subagents parameter. Subagents are useful for context quarantine (keeping the main agent’s context clean) and for providing specialized instructions.

Why use subagents?

Subagents solve the context bloat problem. When agents use tools with large outputs (web search, file reads, database queries), the context window fills up quickly with intermediate results. Subagents isolate this detailed work—the main agent receives only the final result, not the dozens of tool calls that produced it.

When to use subagents:

• ✅ Multi-step tasks that would clutter the main agent’s context
• ✅ Specialized domains that need custom instructions or tools
• ✅ Tasks requiring different model capabilities
• ✅ When you want to keep the main agent focused on high-level coordination

When NOT to use subagents:

• ❌ Simple, single-step tasks
• ❌ When you need to maintain intermediate context
• ❌ When the overhead outweighs benefits

Configuration

subagents should be a list of dictionaries or CompiledSubAgent objects. There are two types:

SubAgent (Dictionary-based)

For most use cases, define subagents as dictionaries with the following fields:

Field	Type	Description
name	str	Required. Unique identifier for the subagent. The main agent uses this name when calling the task() tool. The subagent name becomes metadata for AIMessages and for streaming, which helps to differentiate between agents.
description	str	Required. Description of what this subagent does. Be specific and action-oriented. The main agent uses this to decide when to delegate.
system_prompt	str	Required. Instructions for the subagent. Custom subagents must define their own. Include tool usage guidance and output format requirements. Does not inherit from main agent.
tools	list[Callable]	Required. Tools the subagent can use. Custom subagents specify their own. Keep this minimal and include only what’s needed. Does not inherit from main agent.
model	str | BaseChatModel	Optional. Overrides the main agent’s model. Omit to use the main agent’s model. Inherits from main agent by default. You can pass either a model identifier string like 'openai:gpt-5' (using the 'provider:model' format) or a LangChain chat model object (init_chat_model("gpt-5") or ChatOpenAI(model="gpt-5")).
middleware	list[Middleware]	Optional. Additional middleware for custom behavior, logging, or rate limiting. Does not inherit from main agent.
interrupt_on	dict[str, bool]	Optional. Configure human-in-the-loop for specific tools. Subagent value overrides main agent. Requires checkpointer. Inherits from main agent by default. Subagent value overrides the default.
skills	list[str]	Optional. Skills source paths. When specified, the subagent will load skills from these directories (e.g., ["/skills/research/", "/skills/web-search/"]). This allows subagents to have different skill sets than the main agent. Does not inherit from main agent. Only the general-purpose subagent inherits the main agent’s skills. When a subagent has skills, it runs its own independent SkillsMiddleware instance. Skill state is fully isolated—a subagent’s loaded skills are not visible to the parent, and vice versa.

CLI users: You can also define subagents as AGENTS.md files on disk instead of in code. The name, description, and model fields map to YAML frontmatter, and the markdown body becomes the system_prompt. See Custom subagents for the file format.

CompiledSubAgent

For complex workflows, use a prebuilt LangGraph graph:

Field	Type	Description
name	str	Required. Unique identifier for the subagent. The subagent name becomes metadata for AIMessages and for streaming, which helps to differentiate between agents.
description	str	Required. What this subagent does.
runnable	Runnable	Required. A compiled LangGraph graph (must call .compile() first).

Using SubAgent

import os
from typing import Literal
from tavily import TavilyClient
from deepagents import create_deep_agent

tavily_client = TavilyClient(api_key=os.environ["TAVILY_API_KEY"])

def internet_search(
    query: str,
    max_results: int = 5,
    topic: Literal["general", "news", "finance"] = "general",
    include_raw_content: bool = False,
):
    """Run a web search"""
    return tavily_client.search(
        query,
        max_results=max_results,
        include_raw_content=include_raw_content,
        topic=topic,
    )

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-5.2",  # Optional override, defaults to main agent model
}
subagents = [research_subagent]

agent = create_deep_agent(
    model="claude-sonnet-4-6",
    subagents=subagents
)

Using CompiledSubAgent

For more complex use cases, you can provide your custom subagents. You can create a custom subagent using LangChain’s create_agent or by making a custom LangGraph graph using the graph API.

If you’re creating a custom LangGraph graph, make sure that the graph has a state key called "messages":

from deepagents import create_deep_agent, CompiledSubAgent
from langchain.agents import create_agent

# Create a custom agent graph
custom_graph = create_agent(
    model=your_model,
    tools=specialized_tools,
    prompt="You are a specialized agent for data analysis..."
)

# Use it as a custom subagent
custom_subagent = CompiledSubAgent(
    name="data-analyzer",
    description="Specialized agent for complex data analysis tasks",
    runnable=custom_graph
)

subagents = [custom_subagent]

agent = create_deep_agent(
    model="claude-sonnet-4-6",
    tools=[internet_search],
    system_prompt=research_instructions,
    subagents=subagents
)

Streaming

When streaming tracing information agents’ names are available as lc_agent_name in metadata. When reviewing tracing information, you can use this metadata to differentiate which agent the data came from.

The following example creates a deep agent with the name main-agent and a subagent with the name research-agent:

import os
from typing import Literal
from tavily import TavilyClient
from deepagents import create_deep_agent

tavily_client = TavilyClient(api_key=os.environ["TAVILY_API_KEY"])

def internet_search(
    query: str,
    max_results: int = 5,
    topic: Literal["general", "news", "finance"] = "general",
    include_raw_content: bool = False,
):
    """Run a web search"""
    return tavily_client.search(
        query,
        max_results=max_results,
        include_raw_content=include_raw_content,
        topic=topic,
    )

research_subagent = {
    "name": "research-agent",
    "description": "Used to research more in depth questions",
    "system_prompt": "You are a great researcher",
    "tools": [internet_search],
    "model": "claude-sonnet-4-6",  # Optional override, defaults to main agent model
}
subagents = [research_subagent]

agent = create_deep_agent(
    model="claude-sonnet-4-6",
    subagents=subagents,
    name="main-agent"
)

As you prompt your deepagents, all agent runs executed by a subagent or deep agent will have the agent name in their metadata. In this case the subagent with the name "research-agent", will have {'lc_agent_name': 'research-agent'} in any associated agent run metadata:

Structured output

All subagents support structured output which you can use to validate the subagent’s output.

You can set a desired structured output schema by passing it as the response_format argument to the call to create_agent(). When the model generates the structured data, it’s captured and validated. The structured object itself is not returned to the parent agent. When using structured output with subagents, include the structured data in the ToolMessage.

For more information, see response format.

The general-purpose subagent

In addition to any user-defined subagents, Deep Agents have access to a general-purpose subagent at all times. This subagent:

• Has the same system prompt as the main agent
• Has access to all the same tools
• Uses the same model (unless overridden)
• Inherits skills from the main agent (when skills are configured)

Override the general-purpose subagent

Include a subagent with name="general-purpose" in your subagents list to replace the default. Use this to configure a different model, tools, or system prompt for the general-purpose subagent:

from deepagents import create_deep_agent

# Main agent uses Claude; general-purpose subagent uses GPT
agent = create_deep_agent(
    model="claude-sonnet-4-6",
    tools=[internet_search],
    subagents=[
        {
            "name": "general-purpose",
            "description": "General-purpose agent for research and multi-step tasks",
            "system_prompt": "You are a general-purpose assistant.",
            "tools": [internet_search],
            "model": "openai:gpt-4o",  # Different model for delegated tasks
        },
    ],
)

When you provide a subagent with the general-purpose name, the default general-purpose subagent is not added. Your spec fully replaces it.

When to use it

The general-purpose subagent is ideal for context isolation without specialized behavior. The main agent can delegate a complex multi-step task to this subagent and get a concise result back without bloat from intermediate tool calls.

Example

Instead of the main agent making 10 web searches and filling its context with results, it delegates to the general-purpose subagent: task(name="general-purpose", task="Research quantum computing trends"). The subagent performs all the searches internally and returns only a summary.

Skills inheritance

When configuring skills with create_deep_agent:

• General-purpose subagent: Automatically inherits skills from the main agent
• Custom subagents: Do NOT inherit skills by default—use the skills parameter to give them their own skills

Only subagents configured with skills get a SkillsMiddleware instance—custom subagents without a skills parameter do not. When present, skill state is fully isolated in both directions: the parent’s skills are not visible to the child, and the child’s skills are not propagated back to the parent.

from deepagents import create_deep_agent

# Research subagent with its own skills
research_subagent = {
    "name": "researcher",
    "description": "Research assistant with specialized skills",
    "system_prompt": "You are a researcher.",
    "tools": [web_search],
    "skills": ["/skills/research/", "/skills/web-search/"],  # Subagent-specific skills
}

agent = create_deep_agent(
    model="claude-sonnet-4-6",
    skills=["/skills/main/"],  # Main agent and GP subagent get these
    subagents=[research_subagent],  # Gets only /skills/research/ and /skills/web-search/
)

Best practices

Write clear descriptions

The main agent uses descriptions to decide which subagent to call. Be specific:

✅ Good: "Analyzes financial data and generates investment insights with confidence scores"

❌ Bad: "Does finance stuff"

Keep system prompts detailed

Include specific guidance on how to use tools and format outputs:

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],
}

Minimize tool sets

Only give subagents the tools they need. This improves focus and security:

# ✅ Good: Focused tool set
email_agent = {
    "name": "email-sender",
    "tools": [send_email, validate_email],  # Only email-related
}

# ❌ Bad: Too many tools
email_agent = {
    "name": "email-sender",
    "tools": [send_email, web_search, database_query, file_upload],  # Unfocused
}

Choose models by task

Different models excel at different tasks:

subagents = [
    {
        "name": "contract-reviewer",
        "description": "Reviews legal documents and contracts",
        "system_prompt": "You are an expert legal reviewer...",
        "tools": [read_document, analyze_contract],
        "model": "claude-sonnet-4-6",  # Large context for long documents
    },
    {
        "name": "financial-analyst",
        "description": "Analyzes financial data and market trends",
        "system_prompt": "You are an expert financial analyst...",
        "tools": [get_stock_price, analyze_fundamentals],
        "model": "openai:gpt-5",  # Better for numerical analysis
    },
]

Return concise results

Instruct subagents to return summaries, not raw data:

data_analyst = {
    "system_prompt": """Analyze the data and return:
    1. Key insights (3-5 bullet points)
    2. Overall confidence score
    3. Recommended next actions

    Do NOT include:
    - Raw data
    - Intermediate calculations
    - Detailed tool outputs

    Keep response under 300 words."""
}

Common patterns

Multiple specialized subagents

Create specialized subagents for different domains:

from deepagents import create_deep_agent

subagents = [
    {
        "name": "data-collector",
        "description": "Gathers raw data from various sources",
        "system_prompt": "Collect comprehensive data on the topic",
        "tools": [web_search, api_call, database_query],
    },
    {
        "name": "data-analyzer",
        "description": "Analyzes collected data for insights",
        "system_prompt": "Analyze data and extract key insights",
        "tools": [statistical_analysis],
    },
    {
        "name": "report-writer",
        "description": "Writes polished reports from analysis",
        "system_prompt": "Create professional reports from insights",
        "tools": [format_document],
    },
]

agent = create_deep_agent(
    model="claude-sonnet-4-6",
    system_prompt="You coordinate data analysis and reporting. Use subagents for specialized tasks.",
    subagents=subagents
)

Workflow:

1. Main agent creates high-level plan
2. Delegates data collection to data-collector
3. Passes results to data-analyzer
4. Sends insights to report-writer
5. Compiles final output

Each subagent works with clean context focused only on its task.

Context management

When you invoke a parent agent with runtime context, that context automatically propagates to all subagents. The parent’s full config (including context) is passed through to each subagent invocation internally.

This means tools running inside any subagent can access the same context values you provided to the parent:

from deepagents import create_deep_agent
from langchain.agents import tool
from pydantic import BaseModel

@tool
def get_user_data(query: str, config) -> str:
    """Fetch data for the current user."""
    user_id = config.get("context", {}).get("user_id")
    return f"Data for user {user_id}: {query}"

research_subagent = {
    "name": "researcher",
    "description": "Conducts research for the current user",
    "system_prompt": "You are a research assistant.",
    "tools": [get_user_data],
}

agent = create_deep_agent(
    model="claude-sonnet-4-6",
    subagents=[research_subagent],
    context_schema={"user_id": str, "session_id": str},
)

# Context flows to the researcher subagent and its tools automatically
result = await agent.invoke(
    {"messages": [HumanMessage("Look up my recent activity")]},
    {"context": {"user_id": "user-123", "session_id": "abc"}},
)

Per-subagent context

All subagents receive the same parent context. To pass configuration that is specific to a particular subagent, use namespaced keys: prefix context keys with the subagent’s name:

result = await agent.invoke(
    {"messages": [HumanMessage("Research this and verify the claims")]},
    {
        "context": {
            "user_id": "user-123",                        # shared by all agents
            "researcher:max_depth": 3,                    # only for researcher
            "fact-checker:strict_mode": True,             # only for fact-checker
        }
    },
)

Then read the relevant keys inside tools:

@tool
def verify_claim(claim: str, config) -> str:
    """Verify a factual claim."""
    strict_mode = config.get("context", {}).get("fact-checker:strict_mode", False)
    if strict_mode:
        return strict_verification(claim)
    return basic_verification(claim)

Identifying which subagent called a tool

When the same tool is shared between the parent and multiple subagents, you can use the lc_agent_name metadata (the same value used in streaming) to determine which agent initiated the call:

@tool
def shared_lookup(query: str, config) -> str:
    """Look up information."""
    agent_name = config.get("metadata", {}).get("lc_agent_name")
    if agent_name == "fact-checker":
        return strict_lookup(query)
    return general_lookup(query)

You can combine both patterns—use namespaced context for agent-specific configuration and lc_agent_name metadata for branching tool behavior:

@tool
def flexible_search(query: str, config) -> str:
    """Search with agent-specific settings."""
    agent_name = config.get("metadata", {}).get("lc_agent_name", "unknown")
    ctx = config.get("context", {})

    max_results = ctx.get(f"{agent_name}:max_results", 5)
    include_raw = ctx.get(f"{agent_name}:include_raw", False)

    return perform_search(query, max_results=max_results, include_raw=include_raw)

Troubleshooting

Subagent not being called

Problem: Main agent tries to do work itself instead of delegating.

Solutions:

1. Make descriptions more specific:

   # ✅ Good
   {"name": "research-specialist", "description": "Conducts in-depth research on specific topics using web search. Use when you need detailed information that requires multiple searches."}
   
   # ❌ Bad
   {"name": "helper", "description": "helps with stuff"}
   

2. Instruct main agent to delegate:

   agent = create_deep_agent(
       system_prompt="""...your instructions...
   
       IMPORTANT: For complex tasks, delegate to your subagents using the task() tool.
       This keeps your context clean and improves results.""",
       subagents=[...]
   )
   

Context still getting bloated

Problem: Context fills up despite using subagents.

Solutions:

1. Instruct subagent to return concise results:

   system_prompt="""...
   
   IMPORTANT: Return only the essential summary.
   Do NOT include raw data, intermediate search results, or detailed tool outputs.
   Your response should be under 500 words."""
   

2. Use filesystem for large data:

   system_prompt="""When you gather large amounts of data:
   1. Save raw data to /data/raw_results.txt
   2. Process and analyze the data
   3. Return only the analysis summary
   
   This keeps context clean."""
   

Wrong subagent being selected

Problem: Main agent calls inappropriate subagent for the task.

Solution: Differentiate subagents clearly in descriptions:

subagents = [
    {
        "name": "quick-researcher",
        "description": "For simple, quick research questions that need 1-2 searches. Use when you need basic facts or definitions.",
    },
    {
        "name": "deep-researcher",
        "description": "For complex, in-depth research requiring multiple searches, synthesis, and analysis. Use for comprehensive reports.",
    }
]

---

Edit this page on GitHub or file an issue.

Connect these docs to Claude, VSCode, and more via MCP for real-time answers.