enhanced_agents_guide_2025

Enhanced Agents Guide - Haive Framework

Date: August 7, 2025
Version: 1.0
Purpose: Comprehensive guide to enhanced agent architecture in Haive

🚀 Quick Start

The Haive framework provides three powerful enhanced agents as of August 2025:

1. Enhanced Base Agent - Foundation with lifecycle management
2. SimpleAgentV3 - Dynamic agent with hooks and recompilation
3. ReactAgentV4 - Minimal ReAct pattern with reasoning loops

📋 Architecture Overview

Hierarchy

Workflow (pure orchestration - no LLM)
└── Agent (Workflow + Engine)
    ├── SimpleAgent
    ├── ReactAgent  
    └── MultiAgent

Key Innovation: Engine-Centric Generics

from haive.agents.base.enhanced_agent import Agent
from haive.core.engine.aug_llm import AugLLMConfig

# Agents are generic on their engine type
class MyAgent(Agent[AugLLMConfig]):
    pass

🎯 SimpleAgentV3 - The Workhorse

Basic Usage

from haive.agents.simple.agent_v3 import SimpleAgentV3
from haive.core.engine.aug_llm import AugLLMConfig

# Create agent with convenience fields
agent = SimpleAgentV3(
    name="assistant",
    temperature=0.7,          # Auto-syncs to engine
    max_tokens=1000,         # Auto-syncs to engine
    model_name="gpt-4",      # Auto-syncs to engine
    debug=True               # Default is True!
)

# Execute
result = await agent.arun("Hello, how can you help?")

Dynamic Tool Management

from langchain_core.tools import tool

@tool
def calculator(expression: str) -> str:
    """Calculate mathematical expressions."""
    return str(eval(expression))

# Add tools dynamically
agent = SimpleAgentV3(
    name="math_assistant",
    tools=[calculator],
    force_tool_use=True  # Forces tool usage when available
)

# Tools trigger recompilation automatically
agent.add_tool(another_tool)  # Graph rebuilds!

Hooks System

# Add hooks using decorators
@agent.before_run
def log_input(context):
    print(f"Input: {context.input_data}")

@agent.after_run
def log_output(context):
    print(f"Output: {context.output_data}")

@agent.on_error
def handle_error(context):
    print(f"Error: {context.error}")

# Execute with hook monitoring
result = await agent.arun("Calculate 15 * 23")