Strands Deep Agents

Build sophisticated AI agent systems with planning, sub-agent delegation, and multi-step workflows. Built on the Strands Agents SDK.

What is Deep Agents?

Deep Agents enables AI systems to handle complex, multi-step tasks through:

• 🧠 Strategic Planning - Break down complex tasks into actionable TODOs
• 🤝 Sub-agent Orchestration - Delegate specialized tasks to focused agents
• 📁 Context Management - Efficient file-based operations to keep context lean
• 💾 Session Persistence - Resume work across sessions
• 🔬 Research & Analysis - Multi-perspective investigation and synthesis

Featured Example: DeepSearch - A production-ready research agent demonstrating sub-agent orchestration, parallel research, and intelligent synthesis.

Check the related article for more details.

Installation

# Using UV
uv add strands-deep-agents

# Using pip
pip install strands-deep-agents

Requirements: Python >= 3.12

Quick Start

Basic Agent

from strands_deep_agents import create_deep_agent

agent = create_deep_agent(
    instructions="You are a helpful coding assistant.",
    model="global.anthropic.claude-sonnet-4-5-20250929-v1:0"
)

result = agent("Create a Python calculator module with add, subtract, multiply, divide functions")
print(result)

The agent will automatically:

1. Plan the task using TODOs
2. Create the file with proper structure
3. Add docstrings and examples

Check Task Progress

# View the agent's planned tasks
for todo in agent.state.get("todos", []):
    print(f"[{todo['status']}] {todo['content']}")

Output:

[completed] Plan calculator module structure
[completed] Create calculator.py file
[completed] Add arithmetic functions
[completed] Add docstrings and examples

DeepSearch: Advanced Research Agent

DeepSearch demonstrates the full power of Deep Agents through a production-ready research agent system with intelligent orchestration and parallel research capabilities.

For this example, we can use any internet search tool, for example [Linkup] (https://app.linkup.so/) or tavily tool for web search.

Architecture

DeepSearch uses a three-tier agent architecture:

1. Research Lead Agent - Strategic planning, task decomposition, and synthesis
2. Research Sub-agents - Parallel, focused investigation on specific topics
3. Citations Agent - Post-processing to add proper source references

Key Features

• Intelligent Task Decomposition: Analyzes queries as depth-first (multiple perspectives) or breadth-first (independent sub-questions)
• Parallel Research: Deploys multiple sub-agents simultaneously for efficient information gathering
• Context Management: Sub-agents write findings to files, keeping context lean
• Source Citation: Automated citation agent adds proper references to final reports
• Session Persistence: Resume research across sessions

Running DeepSearch

cd examples/deepsearch

# Basic usage with default prompt
python agent.py

# Custom research prompt
python agent.py -p "Research the impact of AI on healthcare in 2025"

DeepSearch Implementation

from strands_deep_agents import create_deep_agent, SubAgent
from strands_tools import tavily, file_read, file_write

# Research sub-agent - performs focused investigations
research_subagent = SubAgent(
    name="research_subagent",
    description=(
        "Specialized research agent for focused investigations. "
        "Researches specific questions, gathers facts, analyzes sources. "
        "Has access to web search. Writes findings to files."
    ),
    tools=[tavily, file_write],
    prompt="You are a thorough researcher. Gather comprehensive, accurate information..."
)

# Citations agent - adds source references
citations_agent = SubAgent(
    name="citations_agent",
    description="Adds proper citations to research reports.",
    tools=[file_read, file_write],
    prompt="Add citations to research text using provided sources..."
)

# Create the research lead agent
agent = create_deep_agent(
    instructions="""
    You are an expert research lead. Your role:
    1. Analyze the research question
    2. Create a research plan
    3. Deploy research sub-agents for focused investigations
    4. Synthesize findings into a comprehensive report
    5. Call citations agent to add references
    """,
    subagents=[research_subagent, citations_agent],
    tools=[file_read, file_write]
)

# Execute research
result = agent("""
Research AI safety in 2025:
1. Main challenges and concerns
2. Leading organizations and initiatives
Create a comprehensive report with executive summary.
""")

How DeepSearch Works

1. Planning Phase: Lead agent analyzes the query and creates a research plan
2. Parallel Research: Multiple research sub-agents investigate different aspects simultaneously
3. File-Based Context: Sub-agents write findings to ./research_findings_*.md files
4. Synthesis: Lead agent reads all findings and synthesizes a comprehensive report
5. Citations: Citations agent adds proper source references to the final report

Example Output Structure

Research findings written to: research_findings_1.md, research_findings_2.md, ...
Final report: ai_safety_2025_report.md

TODOs:
✅ Analyze research question and create plan
✅ Deploy subagent: AI safety challenges
✅ Deploy subagent: Leading organizations
✅ Deploy subagent: Recent initiatives
✅ Synthesize findings into report
✅ Add citations to report

Core Patterns

Sub-agent Delegation

Deep Agents excel at delegating specialized tasks:

from strands_deep_agents import create_deep_agent

subagents = [
    {
        "name": "researcher",
        "description": "Conducts focused research on specific topics",
        "prompt": "You are a thorough researcher. Gather comprehensive information."
    },
    {
        "name": "analyst",
        "description": "Analyzes data and identifies patterns",
        "prompt": "You are a data analyst. Find insights and patterns."
    }
]

agent = create_deep_agent(
    instructions="You are a research coordinator.",
    subagents=subagents
)

Session Persistence

Resume work across sessions:

from strands.session.file_session_manager import FileSessionManager

session_manager = FileSessionManager(
    session_id="project-xyz",
    storage_dir="./sessions"
)

agent = create_deep_agent(
    instructions="You are a research assistant.",
    session_manager=session_manager
)

# First session
agent("Start researching quantum computing")

# Later - conversation history and TODOs restored
agent("Continue where we left off")

Custom Tools

Extend agents with your own tools:

from strands import tool

@tool
def web_search(query: str) -> str:
    """Search the web for information."""
    # Your implementation
    return search_results

agent = create_deep_agent(
    instructions="You are a research assistant.",
    tools=[web_search]
)

API Reference

create_deep_agent()

Create a deep agent with planning and sub-agent capabilities.

from strands_deep_agents import create_deep_agent

agent = create_deep_agent(
    instructions="System prompt for the agent",
    model=None,                          # Default: Claude Sonnet 4
    subagents=None,                      # List of SubAgent configs
    tools=None,                          # Additional custom tools
    session_manager=None,                # For persistence
    disable_parallel_tool_calling=False, # Force sequential execution
    **kwargs                             # Additional Agent params
)

SubAgent

Define specialized sub-agents:

from strands_deep_agents import SubAgent

subagent = SubAgent(
    name="unique_name",
    description="When to use this agent (helps main agent decide)",
    prompt="System prompt for sub-agent",
    tools=[...],     # Optional: specific tools
    model=model      # Optional: override model
)

Async Support

from strands_deep_agents import async_create_deep_agent

agent = await async_create_deep_agent(
    instructions="You are a research assistant."
)
result = await agent.invoke_async("Your task")

Configuration

Environment Variables

# AWS Bedrock (default provider)
export AWS_REGION=us-east-1
export AWS_ACCESS_KEY_ID=your_key
export AWS_SECRET_ACCESS_KEY=your_secret

# Optional: Skip tool consent prompts
export BYPASS_TOOL_CONSENT=true

Model Options

# Use AWS Bedrock models (default)
agent = create_deep_agent(
    model="global.anthropic.claude-sonnet-4-5-20250929-v1:0"
)

# Custom model configuration
from strands.models import Model

custom_model = Model(
    provider="bedrock",
    name="claude-3-5-sonnet-20241022-v2:0",
    region="us-east-1"
)

agent = create_deep_agent(model=custom_model)

Additional Examples

The examples/ directory contains more patterns:

# Basic agent usage
python examples/basic_usage.py

# Sub-agent delegation patterns
python examples/sub_agents.py

# Session persistence
python examples/session_persistence.py

# DeepSearch - Advanced research agent (recommended)
cd examples/deepsearch
python agent.py

Best Practices

1. Clear Instructions - Be specific about agent roles and expectations
2. Strategic Sub-agents - Use sub-agents for specialized, focused tasks
3. Context Management - Use file operations to keep context lean
4. Session Persistence - Enable for long-running or resumable tasks
5. Prompt for Planning - Encourage agents to create plans before execution

Links

• GitHub: https://github.com/lemopian/strands-deep-agents
• Strands Agents SDK: https://github.com/strands-ai/strands-agents
• Article: Deep Agents Using Strands
• Inspiration: https://github.com/langchain-ai/deepagents

License

MIT License - see LICENSE file for details.

Built with ❤️ by PA