Agent Setup: Your First Agent

Knowledge

At its core, an agent consists of three building blocks: a system prompt that defines its behavior, a model that does the thinking, and tools that give it capabilities. In this section, you'll build all three.

The System Prompt -- Your Agent's DNA

The system prompt is the most important decision in agent development. It determines how your agent behaves, what role it takes on, and what boundaries it respects. A good agent system prompt is fundamentally different from a simple chatbot prompt.

The five elements of an agent system prompt:

1. Role: Who is the agent? ("You are a code review assistant for Python projects.")
2. Capabilities: What can it do? ("You can read files, analyze code, and suggest improvements.")
3. Constraints: What must it not do? ("You never modify code directly. You only suggest changes.")
4. Output format: How does it respond? ("Always respond with a list of findings, sorted by severity.")
5. Reasoning instruction: How does it think? ("Analyze the code step by step. Check structure first, then logic, then style.")

Model Selection -- Which LLM for Which Agent?

Not every model is suited for every agent. The choice depends on three factors:

Rule of thumb: Start with a small model. If the quality isn't sufficient, switch to a larger one. Most agents perform better with Sonnet 5 or GPT-5 mini than you'd expect.

Understand

The First Tool Definition

Tools give your agent capabilities beyond text. A tool is a function that the agent can call -- with a clear description so the LLM knows when and how to use it.

Python example with the Anthropic SDK:

import anthropic

client = anthropic.Anthropic()

# Tool definition: The agent can read files
tools = [
    {
        "name": "read_file",
        "description": "Reads the contents of a file. Use this tool "
                       "when you need to analyze the code or content of a file.",
        "input_schema": {
            "type": "object",
            "properties": {
                "file_path": {
                    "type": "string",
                    "description": "The path to the file, e.g. 'src/main.py'"
                }
            },
            "required": ["file_path"]
        }
    }
]

# Start agent with system prompt and tool
response = client.messages.create(
    model="claude-sonnet-5",
    max_tokens=1024,
    system="You are a code review assistant. Analyze code "
           "step by step and provide concrete improvements.",
    tools=tools,
    messages=[
        {"role": "user", "content": "Review the file src/main.py"}
    ]
)

TypeScript example with the Anthropic SDK:

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic();

// Tool definition
const tools: Anthropic.Tool[] = [
  {
    name: "read_file",
    description:
      "Reads the contents of a file. Use this tool " +
      "when you need to analyze the code or content of a file.",
    input_schema: {
      type: "object" as const,
      properties: {
        file_path: {
          type: "string",
          description: "The path to the file, e.g. 'src/main.py'",
        },
      },
      required: ["file_path"],
    },
  },
];

// Start agent
const response = await client.messages.create({
  model: "claude-sonnet-5",
  max_tokens: 1024,
  system:
    "You are a code review assistant. Analyze code " +
    "step by step and provide concrete improvements.",
  tools,
  messages: [
    { role: "user", content: "Review the file src/main.py" },
  ],
});

Note: The description in the tool is critical. The LLM decides based on this description when to use the tool. A vague description leads to incorrect or missing tool usage.

Apply

Reflect