Originally published on Medium.
In this article we are going to cover 4 commonly confused concepts in agentic AI.
Understanding these distinctions will help you avoid confusing LLM-powered workflows with AI agents.
1. Using an LLM doesn’t make a system an AI Agent.
What makes a system an AI Agent is the LLM’s role in deciding and directing the execution process, not simply whether the system uses an LLM.
To understand why, we are going to distinguish between an agentic workflow and an AI agent.
Both fall under Agentic AI.
The difference is how much control the LLM has over the execution process.
1. Agentic Workflow
An agentic workflow is an AI system where the developer defines the execution path in advance, and the LLM cannot modify or redesign that path.
However, the LLM may make decisions within the execution process, such as classifying user intent or generating a response, but the surrounding code controls how those steps are connected and executed.
Example: “What is the price of an iPhone 15?”
User message
↓
Classify intent ← LLM identifies "get_product_price"
↓
Price node ← Code retrieves the price
↓
Response node ← LLM generates the reply
↓
"The iPhone 15 costs $1000."
2. AI Agent
An AI agent is an AI system where the LLM dynamically decides which steps to take, which tools to use, and whether to continue, based on the task and the results it observes.
Unlike a workflow, the LLM can determine and adjust its own execution path.
Example: “I want the black iPhone, the newer one, if it’s in stock.”
User message
↓
Agent (LLM) decides to search for the product
↓
Calls search_products("black iPhone")
↓
Receives: iPhone 14, iPhone 15, iPhone 15 Pro
↓
Agent decides to check the latest model first
↓
Calls check_stock("iPhone 15 Pro", color="black")
↓
Receives: Out of stock
↓
Agent decides to check another suitable model
↓
Calls check_stock("iPhone 15", color="black")
↓
Receives: 12 in stock
↓
Agent decides it has enough information
↓
Responds to the user
In the above example, the LLM determines which tool to call next based on the user’s request and the results returned by previous tool calls.
2. All agents are agentic AI, but not all agentic AI systems are agents.
The word agentic AI sounds like it means “has an agent in it.”
So people assume any agentic AI system is an agent.
An agentic AI system is a system where an LLM’s decisions influence what the system does, either within steps defined by the developer (a workflow) or by dynamically directing the steps itself (an agent).
There are two common ways to design an agentic AI system:
Agentic Workflow: The developer defines the execution path.
AI Agent: The LLM dynamically decides which actions to take next based on the task and the results of previous actions.
Both can be considered agentic AI systems under this broad definition, although terminology varies across sources.
For example, a LangGraph workflow that uses an LLM to classify a request and routes it through developer-defined steps is an agentic AI system, but not necessarily an AI agent.
A LangGraph agent that uses an LLM to dynamically choose tools, observe their results, and decide what to do next is both an AI agent and an agentic AI system.
3. An agent is a loop, not a single call
An AI agent doesn’t simply call an LLM once and finish.
It repeatedly makes decisions, takes actions, observes the results, and decides what to do next until it considers the task complete.
An agent uses the result of one step to determine its next step.
The loop is what allows an agent to handle tasks that require multiple actions, respond to unexpected results, and correct its approach along the way.
Basic agent loop
Press enter or click to view image in full size
Decide: The LLM looks at the request and everything it knows so far, and picks the next action.
Act: The agent uses a tool to carry out that action.
Observe: The tool’s result is added to the agent’s context as new information.
Decide again: The LLM uses the new information to determine what should happen next. It might call another tool, repeat an earlier action, or conclude that the task is finished.
The loop ends when the LLM gives a final answer instead of calling a tool, or when it hits a limit you set, like a maximum number of turns or an error it can’t recover from.
4. Tool calling alone doesn’t make a system an agent.
Tool calling is a capability that allows an LLM to call an external function or API to perform an action.
What determines whether the system behaves as an agent is what or who controls the decisions surrounding that tool call.
If the developer decides when a tool runs and what happens after, it’s still a workflow.
It becomes an agent when the LLM decides whether to call the tool, which tool to call, and what to do with the result.
Conclusion
In conclusion, understanding the difference between AI agents and agentic workflows comes down to who controls execution.
Top comments (0)