​
Overview

Agents are the core building blocks of evolveRL. Each agent is an autonomous entity that can:
  • Process and respond to messages using an LLM backend
  • Maintain and evolve its own prompt template
  • Track its performance score
  • Save and load its configuration

​
Agent Configuration

Agents are configured using the AgentConfig class: 
from evolverl.agent import Agent, AgentConfig
from evolverl.llm import LLMConfig

# Create agent config
config = AgentConfig(
    llm_config=LLMConfig(
        model_name="gpt-4o-mini",
        model_type="openai"
    ),
    prompt_template="You are a helpful AI assistant...",
    score=0  # Initial score
)

# Create agent
agent = Agent(config)

​
Configuration Options

​
llm_config
LLMConfig
required
Configuration for the LLM backend
​
prompt_template
string
The system prompt template that guides agent behavior
​
score
integer
Performance score (0-1000) from evaluations

​
Using Agents

​
Basic Message Exchange

# Create and configure agent
agent = Agent(AgentConfig(LLMConfig()))
agent.set_default_prompt("You are a helpful AI assistant.")

# Send a message
response = await agent.send_message("What is the capital of France?")
print(response)

​
Saving and Loading

# Save configuration
agent.save_config("agent_config.json")

# Load configuration
new_agent = Agent()
new_agent.load_config("agent_config.json")

​
Agent Types

In the evolution process, there are three main types of agents:

​
1. Variant Agents

These are the agents being evolved. They compete against each other and improve through generations. 
# Create base variant
base_agent = Agent(AgentConfig(llm_config))
base_agent.set_default_prompt("Base prompt template...")

# Create mutations
variants = await evolution.mutate_agents(base_agent, count=5)

​
2. Adversary Agents

Generate challenging test cases to evaluate variant agents. 
# Create adversary
adversary = await evolution.build_adversary(
    domain="mathematics",
    description="Generate tricky math problems"
)

​
3. Judge Agents

Evaluate the performance of variant agents. 
# Create judge
judge = await evolution.build_judge(
    domain="mathematics",
    description="Evaluate solution quality"
)

# Get score
score = await evolution.get_agent_score(chat_history, judge)

​
Best Practices

  1. Prompt Templates: Use clear, structured prompt templates with placeholders for dynamic content
  2. Score Tracking: Regularly update agent scores based on performance
  3. Configuration Management: Save agent configurations after significant improvements
  4. Error Handling: Handle LLM API errors gracefully
  5. Testing: Validate agent behavior across different scenarios