AI Agents with LangChain: Practical Patterns

Beyond Simple Prompts

Most LLM applications start with a simple prompt → response pattern. But real-world tasks often require multiple steps, tool use, and the ability to recover from errors. That's where agents come in.

What Makes an Agent?

An AI agent is an LLM that can:

1. Reason about what steps to take
2. Use tools to interact with external systems
3. Remember context from previous interactions
4. Recover from errors and try alternative approaches

Pattern 1: ReAct Agent

The Reasoning + Acting pattern is the most common agent architecture. The LLM thinks about what to do, takes an action, observes the result, and repeats until the task is complete.

Pattern 2: RAG Agent

Retrieval-Augmented Generation adds a knowledge retrieval step. Instead of relying solely on the model's training data, the agent searches a vector database for relevant context before generating a response.

Pattern 3: Multi-Agent Systems

For complex tasks, multiple specialized agents can collaborate. A "manager" agent delegates subtasks to specialist agents (research, coding, review) and synthesizes their outputs.

Practical Tips

• Start simple: Don't build a multi-agent system when a single chain will do
• Structured output: Use Pydantic models to ensure consistent tool outputs
• Streaming: For long-running agents, stream intermediate steps to keep users informed
• Cost control: Set token limits and implement early stopping for runaway agents
• Testing: Unit test your tools independently before integrating with the agent

Real-World Application

For the insurance GPT project, I built a RAG agent that processes claim documents: it extracts key information, classifies the claim type, checks policy coverage, and generates a preliminary assessment — all in a single agent loop.