Lilly Tech Systems
Build a Multi-Agent Workflow
Build a complete, production-ready multi-agent system using LangGraph from scratch. You will create individual agents with tools, orchestrate them with a supervisor, add human-in-the-loop approval gates, and instrument everything with tracing and monitoring — all in 6 hands-on steps.
What You Will Build
A fully functional multi-agent system where specialized agents collaborate to complete complex tasks. A supervisor agent routes work to a researcher, coder, and analyst, each equipped with their own tools. Humans can approve critical actions before execution.
Specialized Agents
Individual ReAct agents with memory, error handling, and domain-specific tools. Each agent is an expert at one job: research, code execution, or data analysis.
Supervisor Orchestration
A LangGraph StateGraph that routes tasks to the right agent, manages handoffs, and assembles final results from multiple agent outputs.
Human-in-the-Loop
Approval gates, interrupt points, and feedback collection so humans stay in control of critical decisions while agents handle the routine work.
Monitoring & Tracing
LangSmith integration for tracing every agent step, cost tracking per run, and structured error handling with automatic retries.
Tech Stack
Every component is open source or has a generous free tier. Total cost to run: $0 for development, minimal in production.
Python 3.11+
The core language for all agent logic, tool implementations, and orchestration code.
LangGraph
Graph-based agent orchestration framework from LangChain. Handles state management, conditional routing, and agent coordination.
LangChain
Agent abstractions, tool wrappers, and LLM integrations that simplify building individual agents.
OpenAI API
gpt-4o-mini for agent reasoning and tool-calling. Structured outputs for reliable supervisor decisions.
Tavily Search
AI-optimized web search API for the research agent. Free tier includes 1,000 searches per month.
LangSmith
Tracing and monitoring platform for debugging agent workflows. Free tier includes 5,000 traces per month.
Prerequisites
Make sure you have these ready before starting.
Required
- Python 3.11 or higher
- An OpenAI API key (
platform.openai.com) - Basic Python knowledge (functions, classes, async/await)
- A terminal (bash, zsh, PowerShell, or CMD)
Helpful but Not Required
- Experience with LangChain or LLM APIs
- Familiarity with graph-based state machines
- Understanding of the ReAct agent pattern
- A Tavily API key for web search (free tier available)
- A LangSmith account for tracing (free tier available)
Build Steps
Follow these lessons in order. Each step builds on the previous one. By the end, you will have a fully deployable multi-agent workflow.
1. Project Setup
Architecture overview, LangGraph fundamentals, agent patterns, tech stack installation, and project structure. You will have a working development environment by the end.
2. Build a Single Agent
Build a complete ReAct agent with tools, memory, structured output, and error handling. This is the building block for your multi-agent system.
3. Tool Infrastructure
Build a reusable tool library with web search, Python code execution, file I/O, and API integrations. Every agent will draw from this shared toolkit.
4. Multi-Agent Orchestration
Build a supervisor agent that routes tasks to specialized workers using LangGraph StateGraph. Implement handoffs, shared state, and result aggregation.
5. Human-in-the-Loop
Add approval gates, interrupt points, and feedback collection. Humans review critical decisions while agents handle routine work autonomously.
6. Monitoring & Debugging
Integrate LangSmith tracing, add cost tracking per agent run, build structured error handling, and create debugging utilities for complex workflows.
7. Enhancements & Best Practices
Parallel execution, streaming output, deployment strategies, scaling patterns, and a comprehensive FAQ for multi-agent system builders.