Building AI Agents with LangChain: A Guide for Enterprise Devs

Most “AI” apps today are just glorified chatbots. You type, they reply. But an Agent can do things. It can browse the web, query a SQL database, and send Slack messages.

The ReAct Pattern (Reasoning + Acting)

The core of any LangChain agent is the ReAct loop. It allows the LLM to pause generation, use a tool, and then continue based on the output.

1. Thought: “The user wants the stock price of Apple. I don’t know it.”
2. Action: “I will use the [Search Tool] to find ‘AAPL stock price’.”
3. Observation: “The tool returned: $185.40.”
4. Final Answer: “Apple stock is currently trading at $185.40.”

Vector Memory (RAG)

Agents have short attention spans (Context Window). To give them long-term memory, we use Vector Databases (like Pinecone or Qdrant).

When a user asks a question, we don’t send the entire database to GPT-4. We embed the query into a vector, find the 3 most similar documents, and construct a prompt.

Code Example: A Simple SQL Agent

Here is how we implement an agent that can query your Postgres database using natural language:

from langchain.agents import create_sql_agent
from langchain.agents.agent_toolkits import SQLDatabaseToolkit
from langchain.sql_database import SQLDatabase
from langchain.llms import OpenAI

db = SQLDatabase.from_uri("postgresql://user:pass@localhost:5432/crm")
toolkit = SQLDatabaseToolkit(db=db, llm=OpenAI(temperature=0))

agent_executor = create_sql_agent(
    llm=OpenAI(temperature=0),
    toolkit=toolkit,
    verbose=True
)

agent_executor.run("How many customers signed up last week?")

The agent will automatically generate the SQL: SELECT COUNT(*) FROM users WHERE created_at > NOW() - INTERVAL '1 week', execute it, and explain the result.