Agent Workflow vs Agent-like Workflow

When designing an LLM-based application, how to structure the agent-based processing flow is a critical decision. As I’ve been studying frameworks like LangGraph, I wanted to define my approach through two distinct patterns: Agent Workflow and Agent-like Workflow.

In this post, I’ll summarize the concepts, differences, and selection criteria between these two approaches.

Agent Workflow

An Agent Workflow leverages the built-in agent systems provided by LangGraph, such as the create_react_agent() functions.

Key Features

• Built-in reasoning loops (e.g., ReAct)
• Automatic management of tool usage, decision-making, and iterative reasoning
• Provides explicit APIs (create_react_agent)
  • AgentExecutor is used when building agents with LangChain. For reference, see the LangChain-Agent section.
• LLM-driven conversational flow and tool invocation

Example Usage

from ai.tools.get_time import get_current_datetime_tool
from ai.tools.repl import python_repl_tool
from ai.tools.search_web import googlenews_search_tool, tavily_search_tool
from ai.tools.retriever import pdf_retriever_tool

my_tools = [python_repl_tool, tavily_search_tool, googlenews_search_tool, 
            get_current_datetime_tool, pdf_retriever_tool]

llm = ChatOpenAI(
    model=openai_api_model,
    openai_api_base=openai_api_base,
    openai_api_key=openai_api_key
)

agent = create_react_agent(
    model=llm,
    tools=my_tools,
    state_modifier="You are a helpful assistant"
)

def run_agent(state: GraphState) -> GraphState:
    response = agent.invoke({"messages": state["messages"]})
    return {"messages": response["messages"]}

workflow = StateGraph(GraphState)
workflow.add_node("MyAgent", run_agent)

workflow.set_entry_point("MyAgent")
workflow.add_edge("MyAgent", END)

Agent Workflow

Agent-like Workflow

An Agent-like Workflow manually implements agent-like logic by explicitly composing workflow nodes using frameworks like LangGraph. Developers configure individual nodes (e.g., LLM nodes, tool nodes, conditional edges) to replicate agent behaviors. I expect to primarily adopt this approach when building with LangGraph going forward.

Key Features

• No reliance on LangGraph’s official agent APIs
• Explicit tool invocation and conditional logic defined by the developer
• Highly customizable workflow design
• Ideal for complex logic or hybrid human-agent collaboration

Example Usage

from ai.tools.get_time import get_current_datetime_tool
from ai.tools.repl import python_repl_tool
from ai.tools.search_web import googlenews_search_tool, tavily_search_tool
from ai.tools.retriever import pdf_retriever_tool

my_tools = [python_repl_tool, tavily_search_tool, googlenews_search_tool, 
            get_current_datetime_tool, pdf_retriever_tool]

llm_with_tools = ChatOpenAI(
    model=openai_api_model,  # OpenRouter 모델 이름
    openai_api_base=openai_api_base,
    openai_api_key=openai_api_key,
    max_tokens=1000
).bind_tools(tools=my_tools)

tool_node = ToolNode(tools=my_tools)

def run_llm_with_tools(state: GraphState) -> GraphState:
	response = llm_with_tools.invoke(state["messages"])
	return {"messages": [response]}

workflow = StateGraph(GraphState)
workflow.add_node("MyLLM", run_llm_with_tools)
workflow.add_node("tools", tool_node)

workflow.set_entry_point("MyLLM")
workflow.add_conditional_edges("MyLLM", tools_condition)
workflow.add_edge("tools", "MyLLM")

Agent-like Workflow

Choosing the Right Workflow

Scenario	Recommended Workflow
Simple QA agents using predefined tools	Agent Workflow
Complex interactions involving conditional logic and human input	Agent-like Workflow
Workflows needing explicit control over reasoning steps	Agent-like Workflow
Developing LangGraph-based LLM apps	Agent-like Workflow

Conclusion

Using LangGraph’s Agent Workflow is beneficial for rapid development and simpler use cases. However, as your application grows in complexity, LangGraph’s Agent-like Workflow provides superior flexibility and control.

Stay tuned for future posts detailing step-by-step implementations and practical examples of building Agent-like Workflows with LangGraph.