DEV Community

Cover image for Build a Production Multi-Agent System with LangGraph and LaunchDarkly in 20 Minutes
LaunchDarkly profile image Scarlett Attensil
Scarlett Attensil for LaunchDarkly

Posted on • Originally published at launchdarkly.com

Build a Production Multi-Agent System with LangGraph and LaunchDarkly in 20 Minutes

#langgraph #python #launchdarkly #rag

Originally published on the LaunchDarkly Documentation

Overview

Build a working multi-agent system with dynamic configuration in 20 minutes using LangGraph multi-agent workflows, RAG search, and LaunchDarkly AI Configs.

Part 1 of 3 of the series: **Chaos to Clarity: Defensible AI Systems That Deliver on Your Goals**

You've been there: your AI chatbot works great in testing, then production hits and GPT-4 costs spiral out of control. You switch to Claude, but now European users need different privacy rules. Every change means another deploy, more testing, and crossed fingers that nothing breaks.

The teams shipping faster? They control AI behavior dynamically instead of hardcoding everything.

This series shows you how to build LangGraph multi-agent workflows that get their intelligence from RAG search through your business documents, enhanced with MCP tools for live external data, all controlled through LaunchDarkly AI Configs without needing to deploy code changes.

What This Series Covers

  • Part 1 (this post): Build a working multi-agent system with dynamic configuration in 20 minutes
  • Part 2: Add advanced features like segment targeting, MCP tool integration, and cost optimization
  • Part 3: Run production A/B experiments to prove what actually works

By the end, you'll have a system that measures its own performance and adapts based on user data instead of guesswork.

What You'll Build Today

In the next 20 minutes, you'll have a LangGraph multi-agent system with:

  • Supervisor Agent: Orchestrates workflow between specialized agents
  • Security Agent: Detects PII and sensitive information
  • Support Agent: Answers questions using your business documents
  • Dynamic Control: Change models, tools, and behavior through LaunchDarkly without code changes

Prerequisites

You'll need:

  • Python 3.9+ with uv package manager (install uv)
  • LaunchDarkly account (sign up for free)
  • OpenAI API key (required for RAG architecture embeddings)
  • Anthropic API key (required for Claude models) or OpenAI API key (for GPT models)

Step 1: Clone and Configure (2 minutes)

First, let's get everything running locally. We'll explain what each piece does as we build.

# Get the code
git clone https://github.com/launchdarkly-labs/devrel-agents-tutorial
cd agents-demo

# Install dependencies (LangGraph, LaunchDarkly SDK, etc.)
uv sync

# Configure your environment
cp .env.example .env
Enter fullscreen mode Exit fullscreen mode

First, you need to get your LaunchDarkly SDK key by creating a project:

  1. Sign up for LaunchDarkly at app.launchdarkly.com (free account).

If you're a brand new user, after signing up for an account, you'll need to verify your email address. You can skip through the new user onboarding flow after that.

  1. Find projects on the side bar
  2. Create a new project called "multi-agent-chatbot"

  3. Get your SDK key:

    ⚙️ (bottom of sidebar) → Projects → multi-agent-chatbot → ⚙️ (to the right)

    → Environments → Production → ... → SDK key

    this is your LD_SDK_KEY

    Now edit .env with your keys:

LD_SDK_KEY=your-launchdarkly-sdk-key  # From step above
OPENAI_API_KEY=your-openai-key        # Required for RAG embeddings
ANTHROPIC_API_KEY=your-anthropic-key  # Required for Claude models
Enter fullscreen mode Exit fullscreen mode

This sets up a LangGraph application that uses LaunchDarkly to control AI behavior. Think of it like swapping actors, directors, even props mid-performance without stopping the show.

⚠️ Do not check the .env into your source control. Keep those secrets safe!

Step 2: Add Your Business Knowledge (2 minutes)

The system includes a sample reinforcement learning textbook. Replace it with your own documents for your specific domain.

# Option A: Use the sample (AI/ML knowledge)
# Already included: kb/SuttonBarto-IPRL-Book2ndEd.pdf

# Option B: Add your documents
rm kb/*.pdf  # Clear sample
cp /path/to/your-docs/*.pdf kb/
Enter fullscreen mode Exit fullscreen mode

Document types that work well:

  • Legal: Contracts, case law, compliance guidelines
  • Healthcare: Protocols, research papers, care guidelines
  • SaaS: API docs, user guides, troubleshooting manuals
  • E-commerce: Product catalogs, policies, FAQs

Step 3: Initialize Your Knowledge Base (2 minutes)

Turn your documents into searchable RAG knowledge:

# Create vector embeddings for semantic search
uv run python initialize_embeddings.py --force
Enter fullscreen mode Exit fullscreen mode

This builds your RAG (Retrieval-Augmented Generation) foundation using OpenAI's text-embedding model and FAISS vector database. RAG converts documents into vector embeddings that capture semantic meaning rather than just keywords, making search actually understand context.

Step 4: Define Your Tools (3 minutes)

Define the search tools your agents will use.

In the LaunchDarkly app sidebar, click Library in the AI section. On the following screen, click the Tools tab, then Create tool.

Create the RAG vector search tool:

Note: we will be creating a simple search_v1 during part 3 when we learn about experimentation.

Create a tool using the following configuration:

Key:

search_v2
Enter fullscreen mode Exit fullscreen mode

Description:

Semantic search using vector embeddings
Enter fullscreen mode Exit fullscreen mode

Schema:

{
  "properties": {
    "query": {
      "description": "Search query for semantic matching",
      "type": "string"
    },
    "top_k": {
      "description": "Number of results to return",
      "type": "number"
    }
  },
  "additionalProperties": false,
  "required": [
    "query"
  ]
}
Enter fullscreen mode Exit fullscreen mode

When you're done, click Save.

Create the reranking tool:

Back on the Tools section, click Add tool to create a new tool. Add the following properties:

Key:

reranking
Enter fullscreen mode Exit fullscreen mode

Description:

Reorders results by relevance using BM25 algorithm
Enter fullscreen mode Exit fullscreen mode

Schema:

{
  "properties": {
    "query": {
      "description": "Original query for scoring",
      "type": "string"
    },
    "results": {
      "description": "Results to rerank",
      "type": "array"
    }
  },
  "additionalProperties": false,
  "required": [
    "query",
    "results"
  ]
}
Enter fullscreen mode Exit fullscreen mode

When you're done, click Save.

The reranking tool takes search results from search_v2 and reorders them using the BM25 algorithm to improve relevance. This hybrid approach combines semantic search (vector embeddings) with lexical matching (keyword-based scoring), making it especially useful for technical terms, product names, and error codes where exact term matching matters more than conceptual similarity.

🔍 How Your RAG Architecture Works

Your RAG system works in two stages: search_v2 performs semantic similarity search using FAISS by converting queries into the same vector space as your documents (via OpenAI embeddings), while reranking reorders results for maximum relevance. This RAG approach significantly outperforms keyword search by understanding context, so asking "My app is broken" can find troubleshooting guides that mention "application errors" or "system failures."

Step 5: Create Your AI Agents in LaunchDarkly (5 minutes)

Create LaunchDarkly AI Configs to control your LangGraph multi-agent system dynamically. LangGraph is LangChain's framework for building stateful, multi-agent applications that maintain conversation state across agent interactions. Your LangGraph architecture enables sophisticated workflows where agents collaborate and pass context between each other.

Create the Supervisor Agent

  1. In the LaunchDarkly dashboard sidebar, navigate to AI Configs and click Create New
  2. Select 🤖 Agent-based
  3. Name it supervisor-agent
  4. Add this configuration:

variation:

supervisor-basic
Enter fullscreen mode Exit fullscreen mode

Model configuration:

Anthropic
Enter fullscreen mode Exit fullscreen mode 
claude-3-7-sonnet-latest
Enter fullscreen mode Exit fullscreen mode

→ Add parameters
→ Click Custom parameters

{"max_tool_calls":5}
Enter fullscreen mode Exit fullscreen mode

Goal or task:

You are a helpful assistant that can search documentation and research papers. When search results are available, prioritize information from those results over your general knowledge to provide the most accurate and up-to-date responses. Use available tools to search the knowledge base and external research databases to answer questions accurately and comprehensively.
Enter fullscreen mode Exit fullscreen mode

Click Review and save. Now enable your AI Config by switching to the Targeting tab and editing the default rule to serve the variation you just created.

Click Edit on the Default rule, change it to serve your supervisor-basic variation, and save with a note like "Enabling new agent config".

The supervisor agent demonstrates LangGraph orchestration by routing requests based on content analysis rather than rigid rules. LangGraph enables this agent to maintain conversation context and make intelligent routing decisions that adapt to user needs and LaunchDarkly AI Config parameters.

Create the Security Agent

Similarly, create another AI Config called security-agent

variation:

pii-detector
Enter fullscreen mode Exit fullscreen mode

Model configuration:

Anthropic
Enter fullscreen mode Exit fullscreen mode 
claude-3-7-sonnet-latest
Enter fullscreen mode Exit fullscreen mode

Goal or task:

You are a privacy agent that REMOVES PII and formats the input for another process. Analyze the input text and identify any personally identifiable information including: Email addresses, Phone numbers, Social Security Numbers, Names (first, last, full names), Physical addresses, Credit card numbers, Driver's license numbers, Any other sensitive personal data. Respond with: detected: true if any PII was found, false otherwise,types: array of PII types found (e.g., ['email', 'name', 'phone']), redacted: the input text with PII replaced by [REDACTED], keeping the text readable and natural. Examples: Input: 'My email is john@company.com and I need help', Output: detected=true, types=['email'], redacted='My email is [REDACTED] and I need help'. Input: 'I need help with my account',Output: detected=false, types=[], redacted='I need help with my account'. Input: 'My name is Sarah Johnson and my phone is 555-1234', Output: detected=true, types=['name', 'phone'], redacted='My name is [REDACTED] and my phone is [REDACTED]'. Be thorough in your analysis and err on the side of caution when identifying potential PII.
Enter fullscreen mode Exit fullscreen mode

This agent detects PII and provides detailed redaction information, showing exactly what sensitive data was found and how it would be handled for compliance and transparency.

Remember to switch to the Targeting tab and enable this agent the same way we did for the supervisor - edit the default rule to serve your pii-detector variation and save it.

Create the Support Agent

Finally, create support-agent

variation:

rag-search-enhanced
Enter fullscreen mode Exit fullscreen mode

Model configuration:

Anthropic
Enter fullscreen mode Exit fullscreen mode 
claude-3-7-sonnet-latest
Enter fullscreen mode Exit fullscreen mode

Click Attach tools.

select: ✅ reranking ✅ search_v2

Goal or task:

You are a helpful assistant that can search documentation and research papers. When search results are available, prioritize information from those results over your general knowledge to provide the most accurate and up-to-date responses. Use available tools to search the knowledge base and external research databases to answer questions accurately and comprehensively.
Enter fullscreen mode Exit fullscreen mode

This agent combines LangGraph workflow management with your RAG tools. LangGraph enables the agent to chain multiple tool calls together: first using RAG for document retrieval, then semantic reranking, all while maintaining conversation state and handling error recovery gracefully.

Remember to switch to the Targeting tab and enable this agent the same way - edit the default rule to serve your rag-search-enhanced variation and save it.

When you are done, you should have three enabled AI Config Agents.

Step 6: Launch Your System (2 minutes)

Start the system:

# Terminal 1: Start the backend
uv run uvicorn api.main:app --reload --port 8000
Enter fullscreen mode Exit fullscreen mode 
# Terminal 2: Launch the UI  
uv run streamlit run ui/chat_interface.py --server.port 8501
Enter fullscreen mode Exit fullscreen mode

Open http://localhost:8501 in your browser. You should see a clean chat interface.

Step 7: Test Your Multi-Agent System (2 minutes)

Test with these queries:

Basic Knowledge Test:
"What is reinforcement learning?" (if using sample docs)
Or ask about your specific domain: "What's our refund policy?"

PII Detection Test:
"My email is john.doe@example.com and I need help"

Workflow Details show:

  • Which agents are activated
  • What models and tools are being used
  • Text after redaction Watch LangGraph in action: the supervisor agent first routes to the security agent, which detects PII. It then passes control to the support agent, which uses your RAG system for document search. LangGraph maintains state across this multi-agent workflow so that context flows seamlessly between agents.

Step 8: Make Changes Without Deploying Code

Try these experiments in LaunchDarkly:

Switch Models Instantly

Edit your support-agent config:

{
  "model": {"name": "chatgpt-4o-latest"}  // was claude
}
Enter fullscreen mode Exit fullscreen mode

Save and refresh your chat. No code deployment or restart required.

Adjust Tool Usage

Want to limit tool calls? Reduce the limits:

{
  "customParameters": {
    "max_tool_calls": 3  // was 5
  }
}
Enter fullscreen mode Exit fullscreen mode

Change Agent Behavior

Want more thorough searches? Update instructions:

{
  "instructions": "You are a research specialist. Always search multiple times from different angles before answering. Prioritize accuracy over speed."
}
Enter fullscreen mode Exit fullscreen mode

Changes take effect immediately without downtime.

Understanding What You Built

Your LangGraph multi-agent system with RAG includes:

1. LangGraph Orchestration
The supervisor agent uses LangGraph state management to route requests intelligently based on content analysis.

2. Privacy Protection
The supervisor agent uses LangGraph state management to route requests intelligently. This separation allows you to assign a trusted model to the security and supervisor agents and consider on a less-trusted model for the more expensive support agent at a reduced risk of PII exposure.

3. RAG Knowledge System
The support agent combines LangGraph tool chaining with your RAG system for semantic document search and reranking.

4. Runtime Control
LaunchDarkly controls both LangGraph behavior and RAG parameters without code changes.

What's Next?

Your multi-agent system is running with dynamic control and ready for optimization.

In Part 2, we'll add:

  • Geographic-based privacy rules (strict for EU, standard for Other)
  • MCP tools for external data
  • Business tier configurations (free, paid)
  • Cost optimization strategies

In Part 3, we'll run A/B experiments to prove which configurations actually work best with real data.

Try This Now

Experiment with:

  1. Different Instructions: Make agents more helpful, more cautious, or more thorough
  2. Tool Combinations: Add/remove tools to see impact on quality
  3. Model Comparisons: Try different models for different agents
  4. Cost Limits: Find the sweet spot between quality and cost

Every change is instant, measurable, and reversible.

Key Takeaways

  • Multi-agent systems work best when each agent has a specific role
  • Dynamic configuration handles changing requirements better than hardcoding
  • LaunchDarkly AI Configs control and change AI behavior without requiring deployments
  • Start simple and add complexity as you learn what works

Related Resources

Explore the LaunchDarkly MCP Server - enable AI agents to access feature flag configurations, user segments, and experimentation data directly through the Model Context Protocol.

Questions? Issues? Reach out at aiproduct@launchdarkly.com or open an issue in the GitHub repo.

This post was originally published on the LaunchDarkly Documentation site. Follow @LaunchDarkly for more AI and feature flag content!

Top comments (1)

Some comments may only be visible to logged-in visitors. Sign in to view all comments.