Quick Start: Your First Agent in 15 Minutes

Build a minimal conversational agent that uses Claude to answer questions. This is the foundation for more advanced agents.

Goal

Create an agent that:

• Takes user input from terminal
• Sends messages to Claude API
• Returns AI-generated responses
• Runs in a simple loop

Setup

pip install anthropic

python -c "import anthropic; print(anthropic.__version__)"

export ANTHROPIC_API_KEY="your-api-key-here"

set ANTHROPIC_API_KEY=your-api-key-here

echo $ANTHROPIC_API_KEY  # macOS/Linux
echo %ANTHROPIC_API_KEY%  # Windows

mkdir my-first-agent
cd my-first-agent
touch agent.py

my-first-agent/
└── agent.py

Core Concepts

The Agent Loop

Every agent follows this pattern:

# Simplified agent loop
while user_wants_to_chat:
    user_input = get_input()
    response = send_to_llm(user_input)
    display(response)

System Prompts

System prompts define agent behavior and personality:

system_prompt = "You are a helpful research assistant."
# Agent will behave according to this instruction

Message Format

Anthropic API uses structured message format:

messages = [
    {"role": "user", "content": "What is an AI agent?"},
    {"role": "assistant", "content": "An AI agent is..."}
]

Build the Agent

Pattern 1: Initialize API Client

import anthropic
client = anthropic.Anthropic()

Pattern 2: Create Message

response = client.messages.create(
    model="claude-3-5-sonnet-20241022",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello"}]
)

Pattern 3: Extract Response

answer = response.content[0].text
print(answer)

Test Your Agent

python agent.py

You: What is an AI agent?
Agent: [Explains AI agents]

You: How do they differ from chatbots?
Agent: [Explains differences]

You: quit
[Program exits]

Understanding the Flow

graph LR
    A[User Input] --> B[Send to API]
    B --> C[Claude Processes]
    C --> D[Generate Response]
    D --> E[Display to User]
    E --> A

Key Points:

1. Each message is independent (no memory yet)
2. API call happens every turn
3. System prompt shapes responses
4. User controls the loop

Common Issues

What You Built

Current Capabilities:

• Single-turn Q&A
• Claude API integration
• Terminal interface
• System prompt configuration

Missing Features:

• ❌ Conversation memory
• ❌ Tool use (web search, calculator)
• ❌ Conversation persistence
• ❌ Multi-turn context

Key Takeaways

1. Agent Loop: Get input → Send to LLM → Display response → Repeat
2. System Prompts: Define agent behavior and personality
3. API Integration: Anthropic SDK handles communication with Claude
4. Stateless by Default: Each turn is independent without memory

Extension Ideas

Before moving to Core Build, try customizing:

• System Prompt: Change agent personality (formal, casual, expert)
• Model Selection: Try different Claude models (Opus, Sonnet, Haiku)
• Response Length: Adjust max_tokens parameter
• Temperature: Add temperature parameter for creativity control