Agent Patterns

Agent Patterns

Creating Agents with LLMConfig

Always use the LLMConfig context manager. Agents created inside its scope inherit the configuration automatically.

from ag2 import LLMConfig
from ag2.agentchat import AssistantAgent, UserProxyAgent

with LLMConfig(api_type="openai", model="gpt-4o"):
    planner = AssistantAgent(
        name="planner",
        system_message="You are a project planner. Break tasks into steps.",
    )
    reviewer = AssistantAgent(
        name="reviewer",
        system_message="You review plans for completeness and correctness.",
    )

executor = UserProxyAgent(
    name="executor",
    human_input_mode="NEVER",
    code_execution_config={"work_dir": "workspace"},
)

Agents that do not need an LLM (e.g., pure code executors) should be created outside the context manager.

System Message Best Practices

1. Be specific about the role. State what the agent does, not what it is.
2. Define boundaries. Tell the agent what it should NOT do.
3. Keep it short. Aim for 2-4 sentences. Long prompts dilute focus.
4. Use second person. Write "You analyze data" not "This agent analyzes data".

# Good
system_message = (
    "You are a SQL analyst. Write SQL queries to answer user questions. "
    "Only use SELECT statements. Never modify data."
)

# Bad -- too vague
system_message = "You are a helpful assistant."

Agent Naming Conventions

• Use lowercase snake_case: data_analyst, code_reviewer.
• Names must be unique within a group chat.
• Names appear in conversation logs, so make them descriptive.
• Avoid generic names like assistant or agent1.

Tool Registration

Register tools on the agents that should be able to call and execute them:

from ag2.tools import tool

@tool
def search_database(query: str, limit: int = 10) -> str:
    """Search the database and return matching rows."""
    # implementation
    return results

analyst = AssistantAgent(name="analyst", system_message="...")
executor = UserProxyAgent(name="executor", human_input_mode="NEVER")

# Register: analyst decides when to call, executor runs the function
analyst.register_tool(search_database, caller=analyst, executor=executor)

Human Input Modes

• "ALWAYS" -- Ask for human input on every turn.
• "TERMINATE" -- Ask only when the agent wants to terminate.
• "NEVER" -- Fully autonomous, no human input.

# Autonomous executor
executor = UserProxyAgent(name="executor", human_input_mode="NEVER")

# Human-in-the-loop
user = UserProxyAgent(name="user", human_input_mode="TERMINATE")

Termination

Agents stop when a reply contains "TERMINATE". Configure this in the system message:

system_message = (
    "You solve math problems. When you have the final answer, "
    "reply with TERMINATE."
)