What separates an AI agent from agentic AI, where the published definitions disagree, and a five-level autonomy scale a board can use.
The term is much older than large language models. The textbook Artificial Intelligence: A Modern Approach, by Stuart Russell and Peter Norvig, makes the intelligent agent its unifying theme, and its preface defines AI as "the study of agents that receive percepts from the environment and perform actions". The fourth edition has a whole chapter on multiagent decision making. Single agents and systems of many agents were studied for decades before the current wave.
Ranjan Sapkota and colleagues date the widespread adoption of both terms to around 2022. Before then, they write, agent-like systems were built for specific tasks with predefined rules, limited autonomy and little ability to adapt. Large language models changed that, and Anthropic ties the arrival of agents in production to better reasoning, planning and tool use in those models.
OpenAI's practical guide for product teams gives a short definition: "Agents are systems that independently accomplish tasks on your behalf." It names three parts: the model that reasons and decides, the tools it uses to fetch information and act in other systems, and the instructions that set its behaviour and guardrails. The guide also says what falls outside the term: applications that use a language model without letting it control how a workflow runs, such as simple chatbots, single-turn models or sentiment classifiers.
Anthropic draws a finer line inside the same family. In a workflow, the model and its tools follow "predefined code paths" written by engineers. An agent belongs to the "systems where LLMs dynamically direct their own processes and tool usage". Anthropic tells developers to start with the simplest design that works, and warns: "The autonomous nature of agents means higher costs, and the potential for compounding errors."
The review by Sapkota, Konstantinos Roumeliotis and Manoj Karkee in the journal Information Fusion sets the two side by side. In its summary table, AI agents are "autonomous software programs that perform specific tasks" and agentic AI means "systems of multiple AI agents collaborating to achieve complex goals". The authors describe agentic AI as a shift "marked by multi-agent collaboration, dynamic task decomposition, persistent memory, and coordinated autonomy".
The authors illustrate the difference with a house, where the AI agent is a smart thermostat: "This standalone agent receives a user-defined temperature setting and autonomously controls the heating or cooling system to maintain the target temperature." An agentic system in the same house has separate agents for weather forecasts, energy prices, appliance schedules and security, and they share information. A heatwave forecast can then lead the energy agent to cool the house early on solar power, before peak prices.
An OECD report, developed by the OECD.AI Expert Group on Agentic AI, also keeps the terms apart. Its published summary calls the two "closely related, but not interchangeable", and describes agentic AI as several co-ordinated agents that break down tasks, collaborate and pursue complex objectives over extended periods, in less predictable settings and with minimal human supervision.
OpenAI's guide makes a comparable split for developers. It separates single-agent systems, where one model with tools runs in a loop, from multi-agent systems, where work passes between agents through a central manager agent or by agents handing tasks to each other. It tells teams to "maximize a single agent's capabilities first" and to consider more agents when one keeps failing to follow complicated instructions or picks the wrong tools.
Rows drawn from Sapkota et al. (abstract, Tables 1 and 2, section on challenges), OpenAI, Anthropic and OECD.AI.
Anthropic files both kinds of system under one heading, agentic systems. The US National Institute of Standards and Technology (NIST) goes further. Its Center for AI Standards and Innovation, asking industry in January 2026 for evidence on how to secure these systems, wrote:
AI agent systems are capable of planning and taking autonomous actions that impact real-world systems or environments. AI agent systems consist of at least one generative AI model and scaffolding software that equips the model with tools to take a range of discretionary actions. These systems may be more expansive, containing multiple sub-agents with software that orchestrates their interactions. They can be deployed with little to no human oversight. Other terms used to refer to AI agent systems include AI agents and agentic AI.
In NIST's usage, a single agent and a team of sub-agents fall in one category. Even the OECD.AI summary says agentic AI "generally refers" to multi-agent systems, which leaves room for other uses.
When a supplier, an adviser or an internal team uses either term, ask for the architecture behind it: how many agents, which tools each one can call, what each can change outside the system, and where a person must step in. NIST built its scope around the third question. The notice covers systems "capable of taking actions that affect external state, i.e., persistent changes outside of the AI agent system itself", and sets aside chatbots that are not orchestrated to act autonomously.
K. J. Kevin Feng, David McDonald and Amy Zhang at the University of Washington leave the naming question aside and look at autonomy, which they define as "the extent to which an AI agent is designed to operate without user involvement". They argue that "autonomy can instead be a deliberate design decision made by agent developers", set apart from capability: an agent that scores well on benchmarks can still be required to consult its user before each action. Their five levels apply to multi-agent systems too, where the user of an agent can be another agent.
Levels and controls from Feng et al.; the board questions are ours.
OpenAI advises human oversight for actions that are "sensitive, irreversible, or have high stakes" until confidence in the agent grows, and names cancelling orders, authorising large refunds and making payments as examples. Level 4 has its own weakness, in Feng and colleagues' words: "User disengagement can lower the care with which they approve actions and can result in unintended approvals."
Sapkota et al. add a warning for multi-agent systems: "a faulty or hallucinated output from one agent can propagate through the system, compounding inaccuracies and corrupting subsequent decisions". Each agent may hold its own memory, objective and reasoning path, so that "tracing the causal chain of a final decision or failure becomes exceedingly difficult". The authors name the lack of shared, transparent logs across agents as one cause.
It means several AI agents working towards one goal: they split it into tasks, share information and keep going over long periods with limited human supervision. NIST and some other authors use the term for any AI agent.
Generative AI produces content in response to a prompt. Sapkota et al. treat it as the foundation for both agents and agentic AI and rate its autonomy as low, because it waits for a prompt. An AI agent adds tools it can use on its own; agentic AI adds co-ordination between agents so that the system manages a whole process.
Usually not, under the OpenAI and NIST texts cited above. A chatbot moves towards being an agent when it can choose and use tools to finish a task, such as updating a record or issuing a refund. Sapkota et al. do list customer service chatbots among AI agent applications, which shows how loosely the word is used.
The large language model is one component. OpenAI describes an agent as a model plus tools plus instructions. Anthropic calls a model with retrieval, tools and memory the "augmented LLM" and treats it as the basic building block of agentic systems.
The first of four panels at the Davos Super Intelligence Summit 2027, on Tuesday 19 January 2027 at the Grandhotel Belvedere in Davos, is titled "The future of AGI and agentic AI". Our piece on AGI and superintelligence covers the AGI side of that title. The full programme lists all four panels; access is by invitation only.