Agent Fundamentals: Chains vs. Agents

Up until now, you have built Chains. A chain has a fixed path: Input -> Prompt -> Model -> Output. Even if you have 10 steps, the path is hardcoded by you, the developer.

An Agent is different. An agent uses an LLM as a Reasoning Engine to determine which steps to take and in what order.

1. The Decision Loop

1. Observe: Look at the user's query and the tools available.
2. Think: "To solve this, I first need to search the web, then calculate the total."
3. Act: Call the tool.
4. Observe: Look at the tool's output.
5. Repeat: Keep going until the problem is solved.

---

2. When to use an Agent

• Chain: "Translate this text to French." (Always the same step).
• Agent: "Research the top 3 AI trends of 2024 and email me a summary." (Needs dynamic steps: Search $\rightarrow$ Read $\rightarrow$ Summarize $\rightarrow$ Send).

---

3. Visualizing the Control Loop

graph TD
    User[Query] --> LLM[Reasoning Engine]
    LLM -->|Think| Decision{Wait, do I need a tool?}
    Decision -->|Yes| T[Tool Call]
    T --> result[Tool Result]
    result --> LLM
    Decision -->|No| Answer[Final Response]

---

4. The ReAct Pattern

Most LangChain agents follow the ReAct (Reason + Act) pattern.

• Thought: "I need to find the capital of France. I will use the search tool."
• Action: search("capital of France")
• Observation: "Paris"
• Thought: "I have the answer. I will respond to the user."

---

5. Engineering Tip: Latency vs. Flexibility

Agents are "Slower" than chains because they usually require Recursive calls to the LLM. Every "Step" adds 1-2 seconds of waiting. Only use an agent if the path to the answer truly cannot be predicted.

---

Key Takeaways

• Agents use LLMs to dynamically choose their own path.
• Chains are deterministic; Agents are autonomous.
• The ReAct loop is the standard reasoning framework.
• Use Agents for Open-ended tasks where the steps aren't known in advance.