🇰🇷 Korean | 🇺🇸 English

Agent Architecture Practice Lab

This repository explores the evolution of LLM-based Agent Architectures through hands-on experiments
from ReAct-centered first-generation agents to workflow-orchestrated Post-ReAct systems.

---

Common Setup

Initializes the shared environment and tools used throughout all experiments.

0. TEST

• Verifies LLM connectivity and API key configuration
• Establishes a working baseline before running agent experiments

1. Tools

• Defines external tools such as Tavily Search
• Connects tools via the LangChain Tool interface
• Prepares the Action Layer used by agents

---

[Gen 1] Prompt-driven Tool-Using Agents (ReAct-based Agents)

Experiments with first-generation agent patterns where the LLM independently performs the Reason → Act → Observe loop.

1. Zero-Shot ReAct

• The purest form of the ReAct pattern
• No planning phase; reasoning happens on-the-fly
• The LLM owns the entire workflow decision process

2. Conversational ReAct

• ReAct augmented with conversation history (memory)
• Observes behavioral changes when context is accumulated
• Transition from stateless to context-aware reasoning

3. Search-augmented ReAct (Self-Ask with Search)

• Breaks problems into sub-questions (Self-Ask strategy)
• Integrates external search tools for information retrieval tasks

4. ReAct Docstore

• Performs QA against a provided document store
• Demonstrates that even with retrieval, control still resides in the LLM

---

[Gen 2] LLM Orchestrated Systems (Post-ReAct Agents)

Moves beyond LLM-driven control toward system-designed execution, where the workflow is defined externally and the LLM acts as a component.

1. Plan-Then-Execute (Two-Phase Execution Architecture)

• Separates planning and execution into distinct phases
• Introduces structural determinism absent in ReAct

2. State Machine (FSM-based Orchestration)

• Defines explicit states and transitions
• Treats the agent as part of a controlled workflow rather than a free-form conversation

3. Graph Execution (DAG Workflow Engine)

• Introduces DAG-based execution modeling
• Enables non-linear workflows and reusable nodes

4. DAG-Orchestrated Stateful Execution (FSM + Graph)

• Combines FSM state control with DAG execution
• Approximates real-world orchestration architectures

5. Role-Specialized Multi-Agent System

• Splits responsibilities across Planner / Researcher / Builder / Critic / Supervisor roles
• Implements a collaborative multi-agent system

6. Deterministic Guardrail

• Adds validation and guardrail mechanisms
• Ensures outputs are controlled at the system level rather than blindly trusting LLM responses

---

Summary

Gen1: LLM-driven systems where the model decides what to do and when to do it.

Gen2: System-driven orchestration where the workflow is engineered and the LLM plays a bounded role.

This lab demonstrates that building LLM agents is no longer just prompt engineering ---
it is fundamentally a software architecture problem.