What Is an AI Agent? The Definition War That Reveals Our Industry's Confusion
Everyone’s building AI agents. Nobody can agree on what one is.
Ask OpenAI and you’ll get a clean technical definition. Ask Yann LeCun and he’ll tell you LLMs aren’t capable of true agency. Ask your startup founder friend and they’ll call their GPT wrapper an “agent” because it has a system prompt.
The confusion isn’t accidental. The term “AI agent” is load-bearing vocabulary for a $200B+ industry. Different camps are fighting to own its definition.
After surveying academics, industry leaders, and practitioners, I’ve found at least five distinct definitions in active use. Understanding which one someone means determines whether you’re having a productive conversation or talking past each other.
The Academic Foundation: Russell and Norvig (1995)
The original definition predates LLMs by decades. Stuart Russell and Peter Norvig’s Artificial Intelligence: A Modern Approach defines an agent as,
“anything that can be viewed as perceiving its environment through sensors and acting upon that environment through actuators”.
That’s it. Perceive and act.
A thermostat qualifies as an agent under this definition. It perceives temperature. It acts by turning on heat. Russell and Norvig emphasize that an ideal agent is rational. It chooses actions that maximize expected performance based on its perceptions.
Their framework classifies agents into five categories:
| Agent Type | Description | Example |
|---|---|---|
| Simple Reflex | Acts on current perception only | Thermostat |
| Model-Based Reflex | Maintains internal state | Cruise control |
| Goal-Based | Acts to achieve goals | GPS navigation |
| Utility-Based | Optimizes a utility function | Trading algorithm |
| Learning | Improves from experience | Recommendation system |
Why this matters: This definition is technology-agnostic. It doesn’t require LLMs, neural networks, or even computers. It’s about the relationship between a system and its environment. When academics say “agent,” they often mean this broader category.
The LLM Era: Industry Rewrites the Definition
LLMs arrived. The industry needed new vocabulary. The academic definition was too broad. Calling ChatGPT an “agent” felt wrong. It wasn’t doing anything in the world. It was just generating text.
Three definitions emerged.
Andrew Ng’s Practical Framing
Andrew Ng offers a practitioner’s definition:
“An AI agent can be defined as software functioning with AI technologies, which has the ability for active interaction with external systems and is designed to accomplish a specific task”.
The key phrase is active interaction. A chatbot that only responds isn’t an agent in Ng’s view. An agent reaches out. It calls APIs. It modifies files. It changes state in the world.
Ng emphasizes three components:
- AI technologies: Can be LLMs, traditional ML, or symbolic AI
- Active interaction: Distinguishes agents from passive responders
- Specific task accomplishment: Requires integration work, not just prompting
Harrison Chase’s Autonomy Threshold
Harrison Chase, founder of LangChain, adds a crucial qualifier:
“An agent is a system that uses an LLM to decide the control flow of an application”.
But he goes further. A system calling tools based on pre-written if/else logic isn’t an agent. The LLM must make decisions at runtime.
Chase’s refined definition:
“An agent is a system that uses an AI model to get a minimum economically valuable task done without human supervision or intervention”.
That phrase does heavy lifting. “Minimum economically valuable task” excludes demos and toys. It requires that the system accomplish something useful. Autonomously.
OpenAI’s Tripartite Model
OpenAI provides the most implementable definition:
"An Agent is an LLM configured with instructions (system prompt defining behavior), a model selection, and tools (functions or APIs the LLM can invoke)".
%%{init: {"layout": "dagre"}}%%
flowchart LR
Instructions[System Prompt] --> Agent[Agent]
Model[Model Selection] --> Agent
Tools[Functions/APIs] --> Agent
Agent --> Decision{Decision Engine}
Decision --> Action[Tool Calls]
Decision --> Response[Text Output]This makes agent construction concrete. Pick a model. Write instructions. Wire up tools. The agent becomes the decision engine. It chooses which tools to call and in what order.
The Skeptic’s View: LeCun’s World Model Requirement
Not everyone buys the LLM-as-agent framing. Yann LeCun, Meta’s Chief AI Scientist, argues that true intelligent behavior requires world models:
“For intelligent behavior, a system must be capable of anticipating future events and predicting the outcomes of its actions”.
LeCun is critical of pure LLM-based agents. He calls them insufficient without integrated planning. An LLM generates plausible text. It doesn’t reason about consequences.
He advocates for Joint Embedding Predictive Architectures (JEPA). These allow agents to reason about abstract representations. Not raw sensory data. The agent builds an internal model of how the world works. It uses that model to plan.
The implication: If LeCun is right, today’s “agents” are sophisticated autocomplete with tool access. True agents need architectural changes we haven’t built yet.
The Production Reality: Agents vs. Workflows
Here’s where theory meets practice. Anthropic makes a clear distinction in their building patterns documentation:
Workflows follow predetermined paths. Fixed logic. “If X, then Y”.
Agents make dynamic decisions at runtime. Based on context and reasoning.
But Gartner’s August 2025 analysis found that less than 5% of enterprise applications have true AI agents. The vast majority deploy workflows. Marketing departments call them “agentic”.
Why the gap? Workflows are deterministic. Testable. Cost-efficient. Agents are flexible but harder to control. Harder to audit. Harder to predict. Most production systems need reliability over flexibility.
| Characteristic | Workflow | True Agent |
|---|---|---|
| Decision Logic | Predetermined | Dynamic |
| Testability | Deterministic | Stochastic |
| Cost Prediction | Accurate | Variable |
| Error Handling | Explicit branches | Emergent behavior |
| Deployment Status (2025) | 95%+ of “agentic” systems | <5% of systems |
The Consensus Core: What Everyone Agrees On
Despite the chaos, most experts agree that agents must have:
- Autonomy: Operate without continuous human direction
- Goal-orientation: Designed to achieve specific objectives
- Tool integration: Access external APIs, data, and functions
- Reasoning: Decompose problems and plan solutions
- Adaptability: Adjust behavior based on outcomes
The ReAct framework from Google Research formalizes this. It combines “verbal reasoning traces and text actions in an interleaved manner”.
Thought: I need to find the current stock price for AAPL
Action: call_stock_api("AAPL")
Observation: {"price": 187.42, "change": "+1.2%"}
Thought: I have the information, now I can respond
Action: generate_response()This loop distinguishes agents from simple tool-calling systems. Think. Act. Observe. Repeat. The reasoning is explicit. The agent adapts based on observations.
Emerging Patterns (2026)
The field keeps evolving. New categories are emerging.
Ambient Agents: LangChain introduced this concept. Agents that operate continuously in the background. They respond to events. They don’t wait for user prompts.
Agentic Reasoning: Recent academic surveys define this as a paradigm shift. It transforms LLMs from static processors into autonomous agents. They plan. They act. They self-evolve through interaction.
Multi-Agent Systems: Frameworks like CrewAI, AutoGen, and LangGraph enable specialized agents to collaborate. Conversation-based coordination. Role-based workflows. Google Cloud distinguishes these from simple AI assistants. Agents are proactive. Assistants are reactive.
Which Definition Should You Use?
It depends on who you’re talking to.
Use Russell-Norvig when:
- Discussing AI architectures with researchers
- Building systems across modalities (not just text)
- Designing robots or embodied systems
Use LLM-era definitions when:
- Building production software with LLMs
- Scoping engineering work
- Discussing with product teams
Use Chase’s autonomy threshold when:
- Evaluating agent claims vs. workflow reality
- Deciding build vs. buy for agentic capabilities
- Filtering marketing hype
Be skeptical of “agent” claims when:
- There’s no tool use, just prompting
- Human approval is required for every action
- The system follows a fixed decision tree
The Bottom Line
The AI industry has a definition problem. It’s not getting resolved soon. “Agent” means at least five different things. It depends on who’s using it. And what they’re selling.
The practical approach: when someone says “agent,” ask what they mean. Can it call tools? Does it make decisions at runtime? Or follow predetermined logic? Can it accomplish tasks without human intervention?
Those questions cut through the marketing. Whether you’re building, buying, or evaluating AI agents, definitional clarity is the first step.
Navigating the agentic AI space? I’d love to hear which definitions resonate with your work. Reach out on LinkedIn.