Membangun Agen LLM Sederhana

Welcome! In this lesson, we'll dive into LLM agents. An agent is an LLM that can make decisions, take actions, and achieve goals by interacting with its environment.

Think of it as an LLM with a brain and hands. The "brain" is the LLM for reasoning, and the "hands" are tools it can use to perform tasks.

Every LLM agent has key parts working together to enable its autonomous behavior:

• LLM (The Brain): The core language model for reasoning, understanding, and decision-making.
• Memory (State): Where the agent stores past interactions, observations, or thoughts to maintain context.
• Tools (Actions): External functions or APIs the agent can call to perform specific tasks.
• Planner (Reasoning): The LLM's ability to break down a complex goal into smaller, manageable steps.

Agents operate in a continuous cycle, often called the Agentic Loop. This loop allows them to adapt and progress towards a goal:

1. Observe: The agent receives new input (e.g., user query) or checks its environment.
2. Think: The LLM processes observations, plans next steps, decides if a tool is needed, and what to do.
3. Act: The agent executes a chosen tool (if applicable) or generates a direct response.

This loop repeats until the goal is achieved or no further action is required.

You've learned about Function Calling and Tool Use. For an agent, tools are absolutely crucial!

They allow the LLM to go beyond just generating text. Tools give agents the ability to *do* things in the real world, such as:

• Search the web for up-to-date information.
• Access databases or read files.
• Perform calculations or data analysis.
• Send emails, interact with APIs, or control other software.

Imagine an agent needs to find the current weather. It doesn't know this directly, but it's aware it has a 'get_weather' tool.

Here's a simplified view of its internal thought process, demonstrating the 'Think' step leading to an 'Act' step:

To build a simple agent, you primarily instruct the LLM to:

• Understand its Role: Define its persona (e.g., "You are a helpful assistant").
• Know Available Tools: Provide a clear list of tools it can use, including their names, descriptions, and how to call them.
• Decide & Act: Empower it to decide when and how to use those tools based on the user's request.

The LLM's inherent reasoning capability drives the entire process, using tools when its internal knowledge isn't sufficient.

The prompt is absolutely key to an agent's behavior. It acts as the agent's initial programming. You need to clearly define:

• Its Persona: "You are a friendly travel agent helping users plan trips."
• Its Goal: "Your main goal is to assist users in finding flights, hotels, and local attractions."
• Available Tools: Describe each tool, its purpose, and the exact syntax for calling it (e.g., search_flights(destination, date)).
• Decision Process: "Always think step-by-step before deciding on an action or generating a response."

This comprehensive prompt guides the LLM to act as an effective agent.

Let's consider a simple "Trip Planner" agent. This agent has access to tools like get_flights(destination, date), get_hotels(city, dates), and get_attractions(city).

User: "Plan a weekend trip to Paris next month."

The agent would follow its loop:

1. Think: The user wants a trip plan. I need flights, hotels, and attractions for Paris.
2. Act (Tool): Call get_flights("Paris", "next month").
3. Act (Tool): Call get_hotels("Paris", "next month").
4. Act (Tool): Call get_attractions("Paris").
5. Think & Act (Response): Combine all results into a coherent trip plan for the user.

While powerful, simple agents, especially with minimal prompting, have limitations:

• Hallucinations: They might invent tool calls, parameters, or facts.
• Poor Planning: Can struggle with highly complex, multi-step reasoning or recovery from errors.
• Cost & Latency: Each "think" and "act" step typically involves an LLM API call, increasing cost and response time.
• Prompt Sensitivity: Small changes in instructions can sometimes drastically alter agent behavior.

These are common challenges that more advanced agentic patterns and frameworks aim to address.

Great job! You've learned the fundamentals of building simple LLM agents.

• Agents are LLMs that can reason, plan, and act autonomously.
• They are composed of an LLM, memory, and tools.
• Agents operate in a continuous Observe-Think-Act loop.
• Effective prompting is crucial for defining an agent's persona, goals, and available tools.

While powerful, simple agents have limitations. In future lessons, we'll explore more advanced agentic patterns and frameworks to build even more capable applications.