An AI agent is a system in which a language model chooses its own next steps and acts through tools until the task is done or a limit stops it. Agentic AI is a looser label, used for systems built from several agents, for any LLM system with some autonomy, or for the product category as a whole. As of November 2026 we found no standards body or regulator that defines the two terms against each other, so treat the difference as a convention that changes from one author to the next.
For a buyer, the more useful question is narrower: which decisions does the model make at runtime, and what can it change when it makes them? Below: who uses which definition, the six kinds of system that get mixed up, and what the label should change in a purchase. Very little, as it turns out.
Who defines what, and where they conflict
Vendors wrote most of the definitions in circulation, and they disagree on more than wording.
Anthropic’s “Building effective agents” (December 2024) draws its line inside the system. Agents are LLMs directing their own process and tool use; workflows run on code paths a developer fixed in advance. “Agentic systems” is Anthropic’s umbrella for both, so a fixed chain of prompts counts as agentic in that vocabulary without being an agent.
OpenAI’s “A practical guide to building agents” (April 2025) is firmer about what falls outside: “Agents are systems that independently accomplish tasks on your behalf.” Applications that use an LLM without letting it control workflow execution “are not agents”, and the guide names “simple chatbots, single-turn LLMs, or sentiment classifiers” as examples.
IBM treats the terms as parts and whole. Its Think explainers describe an AI agent as a system that performs tasks autonomously by designing workflows with the tools available, and agentic AI as a system that reaches a specific goal with limited supervision and consists of AI agents.
Google Cloud’s AI Agent Ecosystem overview is much looser. It describes AI agents as AI-powered applications that understand and respond to natural-language requests, and a figure caption adds generative models that think and act towards goals. Read literally, the first half of that covers a good many chatbots.
The source that makes the contrast its whole subject is a paper by Sapkota, Roumeliotis and Karkee, “AI Agents vs. Agentic AI” (arXiv 2505.10468, 2025, with a journal reference to Information Fusion). They keep “AI agents” for modular, task-specific automation and “agentic AI” for multi-agent collaboration with dynamic task decomposition, persistent memory and coordinated autonomy. It is the cleanest version of the split, and it is a proposal by three academic authors.
Standards bodies do not settle it either. NIST’s AI Agent Standards Initiative page, updated on 14 August 2026, describes agents as capable of autonomous actions and gives no formal definition. The words also turn up as category names: the Linux Foundation’s Agentic AI Foundation, announced in December 2025, hosts MCP, goose and AGENTS.md, and OWASP publishes a Top 10 for Agentic Applications. Neither name separates agents from agentic AI.
One system, three labels
Take a general example. A support flow classifies each incoming email, fetches the relevant policy through a retrieval call, and has one model call draft the reply. A developer fixed every step.
Under Anthropic’s vocabulary that is an agentic system and a workflow, and not an agent. Under OpenAI’s it is not an agent, because the model never controls execution. IBM’s wording does not fit it well, since no agents sit inside. Sapkota et al. would call it neither.
Now let the model decide which of four tools to call, and how many times, before it drafts. Anthropic, OpenAI and IBM would all call that an agent. Whether it is also agentic AI depends on the author: IBM’s definition can include it, as a goal-directed system built on an agent, and Sapkota’s excludes it, because nothing collaborates with anything.
A vendor could sell either version as an “agentic AI platform” and defend the phrase by pointing at a published definition.
The six systems people mix up
The table is our working vocabulary. It follows Anthropic and OpenAI on where the agent line falls. None of the sources above gives a formal definition of “assistant”, so that row describes common usage and nothing firmer.
| System | What it is | Who picks the next step | Can it change other systems? | General example | First question to test |
|---|---|---|---|---|---|
| LLM | A model that turns input text into output text; no tools or loop of its own | Nobody: one call, one answer | No | One API call that classifies the sentiment of a review | Is the output correct? |
| Chatbot | An LLM or rule set behind a conversation, often with retrieval | The user, turn by turn | Usually not; it reads and answers | Policy questions answered from internal documents | Are answers right and traceable to a source? |
| Assistant | A helper inside a tool or chat that drafts and suggests while a person decides | The user, with the model proposing | Only when the user confirms | Drafting a reply inside a service desk | Do people actually review what it proposes? |
| Workflow | Fixed code paths with LLM steps inside them | The developer, in advance | Yes, along the defined paths | Extract invoice fields, validate them, prepare an ERP entry | Does every path pass its tests? |
| Agent | A model in a loop, choosing tools and reading results until done or stopped | The model, inside limits set in code | Yes, through the tools it holds | Investigating a ticket across three systems and proposing a fix | Does it reach the right end state, and what happens when it does not? |
| Agentic system | Anthropic: either of the two rows above. Sapkota et al.: several coordinated agents | Depends on the author, so ask | Yes, often under several sets of permissions | A planner handing subtasks to specialist agents | Who verifies the final result against the original task? |
AI agent vs LLM
The LLM is one component and the agent is the software around it: the loop, the tools, the state, the limits. Swap the model and you still have the same agent. Remove the loop and the tools and you have an API call. Our AI agent architecture reference goes through those components one by one.
AI agent vs chatbot
The interface decides nothing. A chat window can sit in front of an agent, and an agent can run with no chat window at all, triggered by an incoming invoice. OpenAI’s test is the useful one: does the model control the execution of the workflow?
Two familiar examples sit on the chatbot side. Klarna described its 2024 system as an AI assistant, and the Air Canada ruling concerned a chatbot. Neither is evidence of what autonomous agents do. If retrieval and a good answer solve the problem, a chatbot has fewer paths to test and less it can damage.
AI agent vs AI assistant
This is the weakest line of the three, and products cross it with a settings change. In common usage an assistant leaves the decision with a person, which is also where BaFin’s guidance on ICT risks in the use of AI puts its lower bar: AI in critical or important functions needs more controls than self-service assistants under full human supervision. That guidance is explicitly non-mandatory and sets no supervisory expectations, but the logic carries over. Supervision and function drive the controls.
Our advice: when an assistant has an option to act without confirmation, assess it as an agent for every action that option covers.
What the label changes for a buyer
By itself, nothing legal or technical. The AI Act’s Annex III entries are written around purposes, such as assessing the creditworthiness of natural persons or risk assessment and pricing in life and health insurance. An agent, a workflow and a conventional scoring model doing the same job land in the same entry.
Liability behaves the same way. In Moffatt v. Air Canada (2024 BCCRT 149, summarised by McCarthy Tétrault), the airline was held liable for negligent misrepresentation by its chatbot, and its argument that the bot was a separate entity was rejected. A small-claims tribunal is no EU precedent. Still, the case involved no agent at all, so a company should not expect a less autonomous label to protect it.
The label does change one thing: how sceptical you should be. Gartner forecast in June 2025, as reported by Reuters, that more than 40% of agentic AI projects would be cancelled by the end of 2027 because of costs and unclear value, and it put the number of genuine agent vendors at about 130 out of thousands, using the term agent washing for the gap. That is a prediction and an analyst’s estimate, now more than a year old. It is reason enough to ask what sits behind the word.
What really differs when the system is an agent
Testing changes first. A workflow has a known set of paths, and you can test each one. An agent’s path is decided at runtime, so you judge outcomes across repeated runs. Anthropic’s “Demystifying evals for AI agents” (January 2026) suggests starting with 20 to 50 simple tasks drawn from real failures.
Risk grows with what the model may do unprompted. OWASP’s LLM Top 10 for 2026 lists Excessive Agency as LLM03, and our AI agent security article treats that threat model properly. OpenAI’s guide names two triggers for bringing in a person: exceeding failure thresholds, and actions that are “sensitive, irreversible, or have high stakes”.
Capability is also lower than demos suggest. In the TheAgentCompany benchmark by Xu et al., the best agent finished about 30% of tasks autonomously in a simulated company. That is a benchmark, and its latest version dates from September 2025, so read it as a rough prior.
Then there is the plain question of need. Microsoft’s Agent Framework documentation says that if a function can handle the task, you should write the function, and its orchestration patterns guide advises the lowest level of complexity that meets the requirement reliably. Both come from a company with agent tooling to promote, which makes the advice easier to believe. A multi-agent design adds its own cost on top, and production practice is covered in our AI agent development article.
Six questions that replace the label
- Which decisions does the model make at runtime, and which are fixed in code?
- Which tools can it call, under which credentials, and which of those tools write?
- Which actions stop for a person, and is that enforced in code or only asked for in a prompt?
- What ends a run, apart from the model saying it has finished?
- Can we see a full trace of one run, including every tool call and its result?
- If there are several agents, why? What would break with one?
If the answer to the first question is “nothing”, you are buying a workflow, and that may well be the right purchase. Ask for it to be tested and governed as one.
We would also keep “agentic” out of requirements documents altogether. Write down the decisions the system may take and the systems it may touch. Two teams reading “agentic AI” can picture different architectures. Two teams reading “may create tickets, may not close them” cannot. For what happens once that is settled, see how to build an AI agent, and for deployments by business function, AI agents for business.
Frequently asked questions
What is the difference between AI agents and agentic AI?
An AI agent is one system in which a model chooses its next steps and uses tools. Agentic AI has no settled meaning: IBM uses it for a system made of agents, Anthropic’s “agentic systems” also includes fixed workflows, and Sapkota et al. reserve it for multi-agent collaboration.
Does agentic AI mean multi-agent?
In Sapkota et al.’s taxonomy, yes. In Anthropic’s vocabulary, no, because agentic systems include workflows with no agent inside. If a vendor uses the phrase, ask how many agents the product runs and what each one is allowed to do.
What is the difference between an AI agent and a chatbot?
Control over execution. A chatbot answers within a conversation, and the user drives each turn. An agent decides which tool to call next and can act in other systems. A chat interface can front either one, so the interface tells you nothing.
Is an LLM an AI agent?
No. An LLM produces output from input. It becomes part of an agent when software wraps it in a loop with tools, state and limits. OpenAI’s guide says single-turn LLM applications are not agents.
Scoping the right system with AI Superior
AI Superior is an AI engineering company in Darmstadt, Germany, founded in 2019 by AI scientists. We build LLM applications, chatbots, RAG systems and AI agents for B2B clients, and the first part of that work is often settling which row of the table above a task belongs in. Our AI agent development services cover that decision and the build that follows it.
Sources and further reading
- Anthropic: “Building effective agents”, December 2024.
- Grace, Hadfield, Olivares, De Jonghe (Anthropic): “Demystifying evals for AI agents”, January 2026.
- OpenAI: “A practical guide to building agents”, April 2025.
- Google Cloud: AI Agent Ecosystem overview, undated, read November 2026.
- Gutowska, A. (IBM Think): explainer on AI agents, undated, read November 2026.
- Stryker, C. (IBM Think): explainer on agentic AI, undated, read November 2026.
- Sapkota, Roumeliotis, Karkee: “AI Agents vs. Agentic AI”, arXiv 2505.10468, 2025 (journal reference: Information Fusion).
- NIST CAISI: AI Agent Standards Initiative, page updated 14 August 2026.
- Linux Foundation: announcement of the Agentic AI Foundation, 9 December 2025.
- OWASP: GenAI LLM Top 10 2026; Top 10 for Agentic Applications for 2026.
- Xu et al.: “TheAgentCompany”, arXiv 2412.14161, version of September 2025.
- Microsoft: Agent Framework overview, updated October 2026; “AI Agent Orchestration Patterns”, 12 February 2026.
- BaFin: Guidance on ICT Risks in the Use of AI at Financial Entities, English version 23 January 2026.
- Regulation (EU) 2024/1689 (AI Act), Annex III.
- Moffatt v. Air Canada, 2024 BCCRT 149; McCarthy Tétrault summary, 19 February 2024.
- Reuters via The Star: report on Gartner’s agentic AI forecast, 25 June 2025.
- Klarna: press release on its AI assistant, 27 February 2024.