AI Agent Architecture Design: From ReAct to Plan-and-Execute

What is an AI Agent

An AI Agent is an intelligent system that can autonomously perceive its environment, formulate plans, use tools, and execute actions. Unlike the traditional "one question, one answer" mode, an Agent can perform multi-step reasoning, autonomously decide when to call external tools, and how to combine multiple operations to accomplish complex tasks.

ReAct Architecture

ReAct (Reasoning + Acting) is the most classic Agent architecture, which alternates reasoning and acting:

from langchain.agents import initialize_agent, AgentType
from langchain.tools import Tool

# Define tools
tools = [
    Tool(name="Search", func=search_func, description="Search internet information"),
    Tool(name="Calculator", func=calculator_func, description="Perform mathematical calculations"),
    Tool(name="Database Query", func=db_query, description="Query database")
]

# Create ReAct Agent
agent = initialize_agent(
    tools=tools,
    llm=llm,
    agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION,
    verbose=True
)

# Agent autonomously reasons and executes
result = agent.run("Query sales data for the past week, calculate total sales, and search for industry average growth rate for comparison")

The ReAct workflow: Thought → Action → Observation → Thought... loops until the final answer is obtained.

Plan-and-Execute Architecture

The Plan-and-Execute architecture separates planning and execution, first formulating a complete plan, then executing step by step:

from langchain.experimental.plan_and_execute import (
    PlanAndExecute, load_agent_executor, load_chat_planner
)

planner = load_chat_planner(llm)
executor = load_agent_executor(llm, tools, verbose=True)

agent = PlanAndExecute(
    planner=planner,
    executor=executor,
    verbose=True
)

result = agent.run("Analyze the company's Q3 financial report and write an investment analysis report")

Advantages of Plan-and-Execute:

• Global perspective: Plan first, then execute, avoiding short-sighted decisions
• Traceability: Each step has a clear corresponding plan
• Interruptible and resumable: The execution process can be paused and resumed

ReWOO Architecture

ReWOO (Reasoning WithOut Observation) separates reasoning and tool invocation, reducing the number of LLM calls. It first generates a complete tool invocation plan, then executes all tool calls in batch, and finally generates the final answer uniformly. This approach can reduce API calls by more than 50% compared to ReAct.

Architecture Selection Guide

Agent Design Principles

• Tool atomicity: Each tool does one thing and does it well
• Error handling: Agents need to handle tool invocation failures
• Safety boundaries: Sensitive operations (e.g., deletion, payment) require human confirmation
• Observability: Record every step of the Agent's reasoning and actions for debugging