12.02 — Project Setup¶

Overview¶

This lesson covers the project initialization for the Reflexion Agent. The setup is nearly identical to the Reflection Agent (Section 11), with one key addition: the Tavily API key for web search.

Step 1: Initialize the Project¶

mkdir reflexion-agent
cd reflexion-agent
poetry init    # Accept defaults

Step 2: Install Dependencies¶

poetry add python-dotenv black isort langchain langchain-openai langgraph langchain-tavily

Package	Purpose
`python-dotenv`	Load API keys from `.env` file
`black` / `isort`	Code formatting and import sorting
`langchain`	Core framework — prompts, output parsers, message types
`langchain-openai`	OpenAI integration — `ChatOpenAI` for GPT-4 Turbo
`langgraph`	Graph-based workflow orchestration
`langchain-tavily`	Tavily search engine integration for real-time web search

[!NOTE] This project uses GPT-4 Turbo instead of GPT-3.5 (used in Section 11). GPT-4's stronger reasoning is needed for: (1) self-critique that's specific enough to act on, (2) reliable function calling with complex Pydantic schemas, and (3) coherent article revision across multiple iterations.

Step 3: Configure Environment Variables¶

Create a .env file:

# OpenAI (required — LLM for drafting, critiquing, and revising)
OPENAI_API_KEY=sk-your-key-here

# Tavily (required — web search engine)
TAVILY_API_KEY=tvly-your-key-here

# LangSmith (recommended — tracing and observability)
LANGCHAIN_TRACING_V2=true
LANGCHAIN_PROJECT=reflexion-agent
LANGCHAIN_API_KEY=ls-your-key-here

Variable	Why It's Needed
`OPENAI_API_KEY`	All LLM calls (first responder, revisor) use GPT-4 Turbo
`TAVILY_API_KEY`	The Execute Tools node calls Tavily for real-time web search results
`LANGCHAIN_TRACING_V2`	Enables LangSmith tracing — critical for debugging the multi-step agent
`LANGCHAIN_PROJECT`	Names the project in LangSmith dashboard

[!TIP] Get a Tavily API key at tavily.com. Tavily offers a free tier with enough requests for development and testing.

Step 4: Create Boilerplate and Validate¶

# main.py
from dotenv import load_dotenv
load_dotenv()

if __name__ == "__main__":
    print("Hello Reflexion!")

python main.py
# Output: Hello Reflexion!

Project File Structure¶

reflexion-agent/
├── .env                  ← API keys (not committed)
├── main.py               ← Graph definition, nodes, execution
├── chains.py             ← First responder and revisor chains
├── schemas.py            ← Pydantic models (AnswerQuestion, ReviseAnswer)
├── tool_executor.py      ← Tavily search tool + ToolNode setup
├── pyproject.toml         ← Project configuration
└── poetry.lock            ← Exact dependency versions

[!IMPORTANT] Note the increased file count compared to the Reflection Agent (Section 11, which had just main.py + chains.py). The Reflexion Agent adds schemas.py for Pydantic output models and tool_executor.py for search tool configuration. This separation keeps each file focused and testable.

Summary¶

Step	Action	New vs. Section 11
Poetry init	Same as Section 11	Same
Install deps	Added `langchain-tavily`	New — for web search
API keys	Added `TAVILY_API_KEY`	New — for Tavily search engine
File structure	4 Python files instead of 2	New — `schemas.py` + `tool_executor.py`
LLM model	GPT-4 Turbo (was GPT-3.5)	Changed — needs stronger reasoning