deepagents-subagents

name: deepagents-subagents description: Using SubAgentMiddleware to spawn subagents for task delegation, context isolation, and specialized work in Deep Agents. language: python

deepagents-subagents (Python)

Overview

SubAgentMiddleware enables agents to delegate work to specialized subagents via the task tool. Subagents provide:

• Context isolation: Subagent work doesn't clutter main agent's context
• Specialization: Different tools/prompts for specific tasks
• Token efficiency: Large subtask context compressed into single result
• Parallel execution: Multiple subagents can run concurrently

When to Use Subagents

How It Works

Main agent has task tool → creates fresh subagent → subagent executes autonomously → returns final report to main agent.

Default subagent: "general-purpose" - automatically available with same tools/config as main agent.

Defining Subagents

Dictionary-based Subagent

from deepagents import create_deep_agent
from langchain.tools import tool

@tool
def search_papers(query: str) -> str:
    """Search academic papers."""
    return f"Found 10 papers about {query}"

@tool
def summarize_paper(paper_id: str) -> str:
    """Summarize a research paper."""
    return f"Summary of paper {paper_id}"

agent = create_deep_agent(
    subagents=[
        {
            "name": "research",
            "description": "Research academic papers and provide summaries",
            "system_prompt": "You are a research assistant. Search papers and provide concise summaries.",
            "tools": [search_papers, summarize_paper],
            "model": "claude-sonnet-4-5-20250929",  # Optional: override model
        }
    ]
)

result = agent.invoke({
    "messages": [{"role": "user", "content": "Research recent papers on transformers"}]
})
# Main agent calls: task(agent="research", instruction="Research recent papers on transformers")

CompiledSubAgent (Custom LangGraph)

from deepagents import create_deep_agent, CompiledSubAgent
from langgraph.graph import StateGraph

def create_weather_graph():
    workflow = StateGraph(...)
    # Build custom graph
    return workflow.compile()

weather_graph = create_weather_graph()

weather_subagent = CompiledSubAgent(
    name="weather",
    description="Get weather forecasts for cities",
    runnable=weather_graph
)

agent = create_deep_agent(
    subagents=[weather_subagent]
)

Decision Table: Subagent Patterns

Code Examples

Example 1: Research Subagent

from deepagents import create_deep_agent
from langchain.tools import tool

@tool
def web_search(query: str) -> str:
    """Search the web."""
    return f"Search results for: {query}"

@tool
def analyze_data(data: str) -> str:
    """Analyze data and extract insights."""
    return f"Analysis: {data[:100]}..."

agent = create_deep_agent(
    subagents=[
        {
            "name": "researcher",
            "description": "Conduct web research and compile findings",
            "system_prompt": "Search thoroughly, save results to /research/ directory, return concise summary",
            "tools": [web_search],
        },
        {
            "name": "analyst",
            "description": "Analyze data and provide insights",
            "system_prompt": "Provide data-driven insights with specific numbers",
            "tools": [analyze_data],
        }
    ]
)

result = agent.invoke({
    "messages": [{
        "role": "user",
        "content": "Research market trends for EVs, then analyze the data"
    }]
})
# Main agent: task(agent="researcher", ...) -> task(agent="analyst", ...)

Example 2: Subagent with Human-in-the-Loop

from deepagents import create_deep_agent
from langgraph.checkpoint.memory import MemorySaver

agent = create_deep_agent(
    subagents=[
        {
            "name": "code-deployer",
            "description": "Deploy code to production",
            "system_prompt": "Deploy code safely with all checks",
            "tools": [run_tests, deploy_to_prod],
            "interrupt_on": {"deploy_to_prod": True},  # Require approval
        }
    ],
    checkpointer=MemorySaver()  # Required for interrupts
)

Example 3: Subagent with Custom Skills

from deepagents import create_deep_agent

agent = create_deep_agent(
    skills=["/main-skills/"],  # Main agent skills
    subagents=[
        {
            "name": "python-expert",
            "description": "Python code review and refactoring",
            "system_prompt": "Review Python code for best practices",
            "tools": [read_code, suggest_improvements],
            "skills": ["/python-skills/"],  # Subagent-specific skills
        }
    ]
)
# Note: Custom subagents DON'T inherit main agent's skills by default
# General-purpose subagent DOES inherit main agent's skills

Example 4: Default General-Purpose Subagent

from deepagents import create_deep_agent

agent = create_deep_agent()

# Agent can use default "general-purpose" subagent for context isolation
result = agent.invoke({
    "messages": [{
        "role": "user",
        "content": "Analyze this large dataset and summarize the key findings"
    }]
})
# Agent may call: task(instruction="Analyze dataset and summarize")
# Uses general-purpose subagent with same tools/config as main

Boundaries

What Agents CAN Configure

✅ Subagent name and description ✅ Custom tools for subagents ✅ Different models per subagent ✅ Subagent-specific system prompts ✅ Subagent middleware and skills ✅ Human-in-the-loop for subagent tools

What Agents CANNOT Configure

❌ Change the task tool name ❌ Make subagents stateful (they're ephemeral) ❌ Share state directly between subagents ❌ Remove the default general-purpose subagent ❌ Have subagents call back to main agent

Gotchas

1. Subagents Are Stateless

# ❌ Subagents don't remember previous calls
agent.invoke({"messages": [{"role": "user", "content": "task(agent='research', instruction='Find data')"}]})
agent.invoke({"messages": [{"role": "user", "content": "task(agent='research', instruction='What did you find?')"}]})
# Second call won't remember first call - fresh subagent each time

# ✅ Main agent maintains conversation memory, not subagents

2. Custom Subagents Don't Inherit Skills

# ❌ Subagent won't have main agent's skills
agent = create_deep_agent(
    skills=["/main-skills/"],
    subagents=[{"name": "helper", ...}]  # No skills
)

# ✅ Explicitly provide skills to subagent
agent = create_deep_agent(
    skills=["/main-skills/"],
    subagents=[{
        "name": "helper",
        "skills": ["/helper-skills/"],  # Subagent-specific
        ...
    }]
)

# ✅ General-purpose subagent DOES inherit main skills
# agent.invoke() -> task(instruction="...") uses general-purpose with main skills

3. Subagent Results Are Final

# Subagents return a single final message
# They can't have back-and-forth dialogue with main agent

# ❌ Can't do this:
# Main: "task(agent='research', instruction='Find data')"
# Research: "What topic?"
# Main: "AI"
# Research: "Here's AI data"

# ✅ Provide complete instructions upfront
# Main: "task(agent='research', instruction='Find data on AI, save to /research/, return summary')"

4. Subagent Interrupts Require Main Checkpointer

# ❌ Subagent HITL without checkpointer
agent = create_deep_agent(
    subagents=[{
        "name": "deployer",
        "interrupt_on": {"deploy": True}
    }]
)

# ✅ Checkpointer on main agent, not subagent
agent = create_deep_agent(
    subagents=[{
        "name": "deployer",
        "interrupt_on": {"deploy": True}
    }],
    checkpointer=MemorySaver()  # On main agent
)

Full Documentation

• Subagents Guide
• SubAgent Middleware
• Task Delegation