Understanding Agents

Agents are the core building blocks of Legion. Learn how to create and customize agents for your needs.

What is an Agent?

An agent is an autonomous entity that can:

  • Process incoming messages
  • Make decisions
  • Use tools to perform actions
  • Communicate with other agents

Creating Agents

Legion provides a decorator-based approach to create agents:

from legion import agent, tool

@agent(
    model="openai:gpt-4o-mini",  # Specify the model provider and model
    temperature=0.2,             # Control randomness in responses
    tools=[add_numbers, multiply]  # Bind external tools
)
class MathHelper:
    """An agent that helps with basic arithmetic and string operations.

    I can perform calculations and manipulate text based on your requests.
    You have both external tools for math operations and internal tools for formatting.
    """

    # Agent-specific tools can be defined within the agent class
    @tool
    def format_result(self, number: float, prefix: str = "Result: ") -> str:
        """Format a number with a custom prefix."""
        return f"{prefix}{number:.2f}"

Agent Properties

  • model: The underlying model for decision making (e.g., "openai:gpt-4o-mini", "anthropic:claude-3-opus")
  • temperature: Creativity vs determinism balance (0.0-1.0)
  • tools: External tools the agent can use
  • system_prompt: Custom system prompt (can be dynamic)

Supported Model Providers

Legion supports multiple model providers:

  • OpenAI (e.g., "openai:gpt-4o", "openai:gpt-4o-mini")
  • Anthropic (e.g., "anthropic:claude-3-opus", "anthropic:claude-3-sonnet")
  • Gemini (e.g., "gemini:gemini-pro", "gemini:gemini-ultra")
  • Groq (e.g., "groq:llama3-70b-8192", "groq:mixtral-8x7b-32768")
  • Ollama (e.g., "ollama:llama3", "ollama:mistral")

Using Agents

Once you've created an agent, you can use it to process messages:

# Create an instance of the agent
agent = MathHelper()

# Process a message synchronously
response = agent.process("Can you add 2 and 3?")
print(response.content)

# Process a message asynchronously
response = await agent.aprocess("Can you multiply 4 by 5?")
print(response.content)

Next Steps

Learn more about:

  • Tools - Extend agent capabilities
  • Blocks - Build processing pipelines
  • Teams - Coordinate multiple agents
Quick StartTools