AI agents vs agentic AI: what the terms mean and why it matters
An AI agent is a thing you can buy or build. Agentic is a property that a system has to a degree. Vendors have reasons to blur the two. This page puts the definitions side by side, traces where the words came from, and gives you a twelve-point checklist that we score on five real systems so you can judge how agentic a product is before you pay for it.
Two definitions side by side
| AI agent | Agentic AI | |
|---|---|---|
| Part of speech | Noun: a system | Adjective: a property of a system, by degree |
| Textbook root | "An agent is anything that can be viewed as perceiving its environment through sensors and acting upon that environment through actuators"[1] | No textbook definition; a 2024–25 industry term for systems that act toward goals with limited supervision |
| Engineering definition | "Systems where LLMs dynamically direct their own processes and tool usage", as opposed to workflows, "where LLMs and tools are orchestrated through predefined code paths"[2] | The degree to which a system plans, uses tools, remembers, acts between approvals, observes the result and recovers from failure |
| Vendor usage | A product: "Agents", "Copilot agents", "AI Agent node" | A positioning: "agentic AI is an artificial intelligence system that can accomplish a specific goal with limited supervision"[3] |
| Useful question | What is it allowed to touch, and at which step does a person approve? | How much of the path does the model choose, and how much was designed in advance? |
The textbook definition is wide on purpose: a thermostat is an agent under it. That is why the word did useful work for thirty years in research and did no work at all in marketing until language models made the "decide what to do next" part real. "Agentic" arrived to name the thing that was new: the model choosing the steps. So the honest relationship between the terms is that an AI agent is a system, and agentic is how much of the system's behaviour the model decides at run time rather than the designer at build time. A workflow with one model call is an agent in the textbook sense and barely agentic. A coding tool that reads a bug report, edits files, runs the tests and tries again is highly agentic, and its vendor may or may not call it an agent.
Anthropic's engineering guide, written by a company selling the agentic option, adds a line worth keeping: "we recommend finding the simplest solution possible, and only increasing complexity when needed".[2] C Our explainer on autonomous AI agents covers what these systems finish without a human and where they fail, with graded evidence; this page is only about the words and how to see through them.
A short history of the terms
1995 onward. "Intelligent agent" is the organising concept of the standard AI textbook, Russell and Norvig's Artificial Intelligence: A Modern Approach, which classifies agents by how they map perceptions to actions, from simple reflex agents to learning agents.[1] B The field used the word for software that acted on a person's behalf: mail filters, trading bots, game characters.
2022–2023. Chat assistants arrive. They answer; they do not act. The research community starts wiring them to tools and loops (plan, call a tool, read the result, repeat), and "agent" moves from the textbook to the engineering blog. The benchmark papers of 2023–24 define the thing by its tasks: an assistant that must browse, use tools and reason to answer, or a tool-calling system that must follow policy in a customer-service conversation.[4] [5] B
2024–2025. Vendors adopt "agentic". It describes a capability rather than a product, so it attaches to anything: agentic workflows, agentic search, agentic commerce. The search data show the shift. On Google Trends, "agentic ai" increased its share of our future-skills term group by 4.3 percentage points over the twelve months to November 2026, and the phrase "ai agents" grew its share of the agent-productivity group by 10.6 points (DataForSEO Trends, retrieved 2026-11-09). Meanwhile the definitional comparison on this page, "ai agents vs agentic ai", draws about 1,300 US searches a month at a cost per click of USD 14 and fell 32% year on year, while "ai agents explanation" (2,400 a month) rose 247%. People have stopped asking which word is right and started asking what the thing does, which is the right question.
The agentic checklist
Six dimensions, each scored 0, 1 or 2. The total is out of 12, but the first dimension is a gate: a system with 0 on planning is automation, however high its total. The dimensions are ours; there is no industry standard, and anyone selling you a score should say the same.
| Dimension | 0 | 1 | 2 |
|---|---|---|---|
| Planning: who chooses the next step | The designer; the path is fixed | The model chooses within a designed menu or branch | The model decomposes the goal and sequences steps itself |
| Tool use: can it act outside the conversation | No; text in, text out | Calls a fixed set of tools | Chooses among many tools and can be given new ones |
| Memory: does it carry state | None beyond the current turn | Within a session or run | Across sessions, with retrieval |
| Autonomy between approvals: how much it does before a person confirms | A person approves every action | Acts freely on read and draft; stops at send, pay, delete | Configurable up to full runs within a budget |
| Environment feedback: does it observe results and adapt | Fires and forgets | Reads a result and branches on it | Loops: act, observe, revise the plan |
| Recovery: what happens on failure | Stops or repeats the same error | Fixed retries | Diagnoses, tries an alternative, escalates with a reason |
Two readings of the sheet matter more than the total. Planning tells you whether you are buying an agent at all. Autonomy between approvals tells you what the thing can cost you, in money and in error, before anyone looks.
Marketing patterns to watch
- The adjective doing the noun's work. "Agentic" on a page that describes fixed steps with a model filling in one of them. Ask which step the model chooses. If none, it is a workflow, which may be exactly what you want.
- "Autonomous" without a permissions list. A system that cannot be run draft-only has no safe first month. The word tells you nothing; the scopes do.
- A single success rate. A demo is one run; agents are inconsistent across runs. The τ-bench authors introduced pass^k, the chance of succeeding on all k attempts, because a single rate hides it; state-of-the-art agents in their retail domain succeeded on fewer than half of tasks and pass^8 fell below 25%.[5] B Ask for repeat counts on your tasks.
- Analyst forecasts as product facts. "A third of enterprise applications will include agentic AI by 2028" is a Gartner prediction read from a press-release headline, and the same release predicts over 40% of agentic projects cancelled by end-2027; both are grade C in our evidence base.[6] C Neither tells you whether a given product works.
- Hours saved. A vendor case study with hours saved is a claim. The randomised and field evidence on AI productivity shows gains that depend on the task and the user, from large for novices on narrow tasks to a 19% slowdown for experienced developers using early-2025 agentic coding tools who believed they were faster.[7] B A number without a task and a population is not evidence.
Five systems scored on the checklist
Our reading of public documentation on 2026-11-09. The scores describe how agentic each system is, not how good; a 4 that does the job is better than an 11 that does not. The sixth dimension is where the documentation is thinnest and the scores are most uncertain.
| System | Planning | Tools | Memory | Autonomy | Feedback | Recovery | Total | Reading |
|---|---|---|---|---|---|---|---|---|
| A chat assistant without tools (any vendor's basic tier) | 0 | 0 | 1 | 0 | 0 | 0 | 1 | An assistant. Agentic in no useful sense; also the safest thing on the list |
| A standard Zapier Zap (trigger, filter, actions) | 0 | 2 | 0 | 2 | 1 | 1 | 6 | Automation, not an agent: the path is fixed, it acts without approval, which is the point. Fails the planning gate |
| Zapier Agents (natural-language instructions, acts across the connector library)[8] | 1–2 | 2 | 1 | 1–2 | 1 | 1 | 7–9 | An agent product on top of a workflow platform; how agentic depends on how open the instructions are |
| n8n AI Agent node (model chooses among connected tools, with a memory node, inside a workflow)[9] | 2 | 2 | 1–2 | 2 (designer sets gates) | 1–2 | 1 | 9–11 | Highly agentic within the boundary the workflow designer draws; billing per run keeps loops from costing extra on the platform |
| An agentic coding tool in a terminal (as studied by METR, early 2025)[7] | 2 | 2 | 1 | 2 (configurable) | 2 | 2 | 11 | The most agentic consumer-available category; also the one with a randomised study showing a slowdown for experts |
Three observations. The two systems most often sold as "agentic" sit at opposite ends: the Zap scores high on autonomy and zero on planning, the coding tool scores high on everything. The checklist separates them where the marketing does not. Second, the agent products with the best evidence of benefit in the field studies are level-1 and level-2 systems on autonomy, where a person still sends the output.[7] Third, score ranges are honest: a product's agenticness depends on how you configure it, and the configuration, not the brochure, is what you should ask to see. Our comparison of the best AI agents for work applies these questions to named products and what each is allowed to touch.
Why the distinction matters for a buyer
Because the two words carry different risks and different bills. An agent with high autonomy and low planning (automation) fails predictably, in a place the designer can test. An agent with high planning fails in a place nobody designed, because it chose the path. The first needs testing; the second needs evaluation on repeated runs, budgets per run, and gates before send, pay and delete. The human-factors literature adds the third risk, which belongs to the person: people with an automated aid make more errors when the aid is wrong than people without one, and the effect is documented across 74 studies with training, display design and accountability as the mitigators.[10] [11] A The more agentic the system, the more that finding applies to whoever reviews it.
In our Future Demand Radar, "AI agents and delegated work" scores 59.7 out of 100 in the baseline scenario (p10–p90 band 55.2–64.1) and is classed "build now"; under the conservative scenario A it scores 56.0 and stays "build now", under the accelerated scenario C 61.7. Our baseline scenario expects search demand for the category to reach about 1.7 times today's level by 2028 (band 1.5–2.0). Forecasts are model outputs with stated uncertainty, not facts. The vocabulary is likely to keep moving; the checklist is built so that it does not have to.
Glossary
- Agent (textbook). Anything that perceives its environment and acts on it.[1]
- Agent (engineering). A system in which a model directs its own steps and tool use.[2]
- Agentic. The degree to which a system's behaviour is chosen by the model at run time rather than by the designer at build time.
- Workflow. Models and tools orchestrated through predefined code paths.[2]
- Tool call. A structured request from the model to run an external action and return the result.
- Gate. A rule that stops the system for human approval before specified actions.
- pass^k. The probability of succeeding on all k repeated attempts at a task; a measure of reliability.[5]
What to do next
Score the product you are considering on the six dimensions from its documentation, then ask the vendor for the two things the sheet cannot show: the permission scopes and the repeat-run success rate on tasks like yours. The Tool Finder asks the same questions about your work, your data constraints and how much you want to delegate, and returns a shortlist with dated prices. Definitions move every quarter; the Productivity Signal briefing carries a glossary update when they do.
Frequently asked questions
What is the difference between an AI agent and agentic AI?
An AI agent is a system; agentic is a property that system has to a degree. In the textbook sense an agent is anything that perceives and acts,[1] and in the engineering sense it is a system where a model directs its own steps and tool use.[2] C Agentic describes how much of the path the model chooses at run time rather than the designer at build time. A fixed workflow with one model call is barely agentic; a coding tool that plans, edits, tests and retries is highly agentic.
Is agentic AI just a marketing term?
Mostly, and that is not a criticism of the word so much as of how it is used. It names something real, the model choosing steps, but it attaches to anything because it is an adjective. The test is the planning dimension of the checklist: if the designer fixed every step, the system is automation, whatever the page says.
How can I tell how agentic a product really is?
Score it on six dimensions from the documentation: planning, tool use, memory, autonomy between approvals, environment feedback and recovery, each 0 to 2. Then ask for the two things documentation does not show, the permission scopes and the success rate across repeated runs on your tasks. Single success rates hide inconsistency; benchmark authors use pass^k for that reason.[5] B
Is a Zapier Zap an AI agent?
No. A standard Zap scores high on autonomy, it acts without anyone approving each run, and zero on planning, because the designer fixed the path. That makes it automation, which is often the better buy: it fails predictably in places you can test. Zapier's separate Agents product, which takes natural-language instructions and acts across the connector library, is an agent product.[8]
Does more agentic mean better?
No. The field evidence of benefit is strongest for systems where a person still sends the output, and the one randomised study of highly agentic coding tools on real work found experienced developers 19% slower while believing they were faster.[7] B More agentic means more of the path chosen by the model, which means more evaluation, tighter budgets and gates before send, pay and delete. Pick the lowest level that does the job.
Sources and evidence
- Artificial Intelligence: A Modern Approach (4th ed.), chapter 2, Intelligent Agents — Russell, Stuart; Norvig, Peter (2020), report B aima.cs.berkeley.edu/ (retrieved 2026-11-09)
- Building effective agents — Anthropic (2024), vendor page C www.anthropic.com/research/building-effective-agents (retrieved 2026-11-09)
- What is agentic AI? — IBM (2025), vendor page C www.ibm.com/think/topics/agentic-ai (retrieved 2026-11-09)
- GAIA: a benchmark for General AI Assistants — Mialon, Grégoire; Fourrier, Clémentine; Swift, Craig; Wolf, Thomas; LeCun, Yann; Scialom, Thomas (2023), report B doi.org/10.48550/arXiv.2311.12983 (retrieved 2026-11-09)
- τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains — Yao, Shunyu; Shinn, Noah; Razavi, Pedram; Narasimhan, Karthik (2024), report B doi.org/10.48550/arXiv.2406.12045 (retrieved 2026-11-09)
- Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027 (press release; headline verified via search only, page returned 403) — Gartner (2025), report C www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts (retrieved 2026-11-09)
- Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity — Becker, Joel; Rush, Nate; Barnes, Elizabeth; Rein, David (METR) (2025), RCT B doi.org/10.48550/arXiv.2507.09089 (retrieved 2026-11-09)
- Zapier Agents product page — Zapier (2026), vendor page A zapier.com/agents (retrieved 2026-11-09)
- n8n pricing (plan definitions; AI agent features) — n8n GmbH (2026), vendor page A n8n.io/pricing/ (retrieved 2026-11-09)
- Does automation bias decision-making? — Skitka, Linda J.; Mosier, Kathleen L.; Burdick, Mark (1999), RCT B doi.org/10.1006/ijhc.1999.0252 (retrieved 2026-11-09)
- Automation bias: a systematic review of frequency, effect mediators, and mitigators — Goddard, Kate; Roudsari, Abdul; Wyatt, Jeremy C. (2012), systematic review A doi.org/10.1136/amiajnl-2011-000089 (retrieved 2026-11-09)