A chatbot answers you in a chat window and stops. An AI agent works through a task on its own, using tools such as a browser or your email and deciding each next step until the job is done. An agentic workflow sits in between: an AI handles several steps, such as searching, drafting and checking its own work, along a route that a person or a program has mostly laid out. The difference that matters is who decides the steps, and what the AI is allowed to touch.
- Chatbot: you ask, it answers, you act. Nothing happens outside the chat.
- Agentic workflow: a set series of steps, with an AI doing some of them. Predictable, and easier to check.
- AI agent: you give a goal; the AI picks the steps and uses tools to carry them out. Flexible, and riskier.
- It is a sliding scale, not three boxes. The more the AI decides and can touch, the more limits it needs.
| Chatbot | Agentic workflow | AI agent | |
|---|---|---|---|
| What you give it | A question or request. | An input to a process someone built. | A goal. |
| Who decides the steps | You do, between replies. | Mostly the person or code that built the process. | The AI. |
| Takes actions? | No: it writes, you act. | Some, at set points. | Yes, with whatever tools it is given. |
| Example | “Which flights go to Lisbon on Friday?” | Every new support email is sorted, a reply drafted, then checked against policy. | “Book me a morning flight to Lisbon on Friday.” |
| Main risk | A wrong answer you act on. | A mistake repeated at every run. | A wrong action that has already happened. |
A chatbot is a friend on the phone who tells you how to book a train. An agentic workflow is a recipe card: the steps are fixed, and the AI is a helper doing some of them. An AI agent is an assistant you have handed your keys to: you say where you want to go, and it decides how to get there.
What is a chatbot?
The plain point: a program you talk to. It replies, and the rest is up to you.
IBM, the computer company, keeps it short: “A chatbot is a software application that communicates with people through text or voice.”1 Older chatbots followed scripts; today’s AI chatbots, such as ChatGPT, Claude and Gemini, write each reply fresh with a large language model. IBM draws that line between “scripted conversation flows” and chatbots that use large language models.2
A plain chatbot does not act in your other apps. If its advice is wrong, you can still ignore it.
What is an AI agent?
The plain point: an AI that chooses its own steps and uses tools to carry them out.
Anthropic, the maker of Claude, defines agents as “systems where LLMs dynamically direct their own processes and tool usage”.3 The tools are how it acts. With tool use, the model asks a program to run a job, such as a web search, and gets the result back; Anthropic says Claude “determines when to call a tool based on the user’s request and the tool’s description.”4 An agent chains many of those calls towards a goal.
Ask a chatbot which flights go to Lisbon and you get a list. Ask an agent to book one and it may open a travel site and fill in your details. More on AI agent meaning.
What is an agentic workflow?
The plain point: a multi-step process where an AI does real work at several steps, but the route is mostly set in advance.
Anthropic’s guide separates the two ideas: “Workflows are systems where LLMs and tools are orchestrated through predefined code paths.” It calls both kinds agentic systems: “At Anthropic, we categorize all these variations as agentic systems”.3 Its building block is “an LLM enhanced with augmentations such as retrieval, tools, and memory”, and its workflow patterns include prompt chaining, routing, parallelization, orchestrator-workers and evaluator-optimizer. The simplest, prompt chaining, “decomposes a task into a sequence of steps, where each LLM call processes the output of the previous one.”3
The phrase is also used for patterns where the AI loops over its own work. A DeepLearning.AI course taught by the AI educator Andrew Ng lists four: “Reflection,” “Tool Use,” “Planning,” and “Multi-Agent”, describing reflection as “AI critiques its own work and iterates to improve quality”.5 The phrase has no single fixed definition. In this site’s reading, a workflow is “agentic” when the AI does more than fill one blank: it searches, drafts, checks or revises inside a process someone designed.
Is it a sliding scale or three separate things?
The plain point: a sliding scale. The labels mark points along it.
Andrew Ng argued for the adjective “agentic” for this reason: “we can acknowledge that there are different degrees to which systems can be agentic.”6 Princeton researchers describe the signs of moving up the scale: systems that “require less user supervision are more agentic”, as are those “whose control flow is driven by an LLM, and hence dynamic”.7
So a chatbot with a search button, a fixed pipeline with an AI step and an assistant booking your travel are agentic to different degrees. Ask the same two questions of each: who picks the next step, and what can it reach?
Which one do you need?
The plain point: the simplest one that does the job.
Anthropic’s advice to builders is to find “the simplest solution possible, and only increasing complexity when needed.”8 When more is needed, it splits the choice this way: “workflows offer predictability and consistency for well-defined tasks,” “whereas agents are the better option when flexibility and model-driven decision-making are needed at scale.”3
- A question or a draft: a chatbot.
- The same job, many times, the same way: a workflow, because each run can be checked against the last.
- A job whose steps cannot be known in advance: an agent, with tight limits.
That list is this site’s reading.
Why does the risk go up along the scale?
The plain point: advice can be ignored; an action has already happened.
Anthropic warns that “The autonomous nature of agents means higher costs, and the potential for compounding errors.”9 Anything that reads web pages or emails can also meet hidden instructions, called prompt injection; OWASP, an open security community, puts it first on its list of risks for AI applications.10 The protection is what the AI can touch: an agent that cannot pay cannot be tricked into paying. Permission tiers for agents shows how to set that.
Ask two questions of any AI tool: who decides the next step, and what can it touch? A chatbot decides nothing and touches nothing; an agent decides a lot, so give it only the keys the job needs.
Questions people ask
Is a chatbot that can search the web an agent?
Not by Anthropic’s line. A tool on its own does not make an agent; what matters is whether the AI directs its own process (source above). In this site’s reading, a chatbot that runs one search per question is at the low end of the scale.
Is an agentic workflow the same as ordinary automation?
Not quite. In this site’s reading, ordinary automation runs a fixed rule (“every Monday, send this report”) and never decides anything new, while an agentic workflow has an AI doing judgement-type work at some steps, such as drafting or checking its own output.
Do I need an AI agent?
Often not. Anthropic recommends the simplest solution that works (source above); a single question or a fixed set of steps is easier to check.
Are AI agents safe to use?
They can be, with limits. Anthropic recommends testing in a walled-off setting with guardrails (source above); give an agent only the access the task needs and keep approval for anything you cannot undo.
Where to go next
- “Give me each word on its own.” Read chatbot, AI agent and agentic AI.
- “I want to try an agent safely.” Start with AI agents for beginners.
- “I am setting one up.” Read agent guardrails and how to prompt AI agents.
- “Should I use several agents?” See multi-agent: when it helps.
- “How do tools plug in?” See tool use and MCP.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What does the page say is the difference that matters most between a chatbot, an agentic workflow and an AI agent?
- Which company made the AI
- Who decides the steps, and what the AI can touch
- How long the replies are
Show the answer
Answer: Who decides the steps, and what the AI can touch. The page says the real difference is who picks the next step and what the AI is allowed to reach.
2. In an agentic workflow, who mostly sets the route?
- The person or program that designed the process
- The AI, which picks every step freely
- The reader, who approves each reply in a chat
Show the answer
Answer: The person or program that designed the process. The page describes an agentic workflow as AI doing work at several steps along a route mostly laid out in advance.
3. Why does the risk go up as you move from a chatbot towards an agent?
- Agents always use older models
- Agents cannot read web pages
- An agent takes actions, and a wrong action has already happened
Show the answer
Answer: An agent takes actions, and a wrong action has already happened. The page says advice can be ignored but an action cannot, and errors can compound over many steps.
- 1IBM, What Is a Chatbot?, updated 16 Jul 2026, read at source 1 Oct 2026 (as cited on this site’s chatbot meaning page): “A chatbot is a software application that communicates with people through text or voice.”↑
- 2IBM, What Is a Chatbot?, updated 16 Jul 2026, read at source 1 Oct 2026 (as cited on this site’s chatbot meaning page): “Traditional chatbots follow predefined rules, decision trees and scripted conversation flows. AI chatbots use technologies such as natural language processing (NLP), generative AI and large language models (LLMs).”↑
- 3Anthropic, Building effective agents, 19 Dec 2024, read at source 19 Oct 2026“Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage”. ↑a“Workflows are systems where LLMs and tools are orchestrated through predefined code paths.” … “At Anthropic, we categorize all these variations as agentic systems”. ↑b“The basic building block of agentic systems is an LLM enhanced with augmentations such as retrieval, tools, and memory.” … “Prompt chaining decomposes a task into a sequence of steps, where each LLM call processes the output of the previous one.” ↑c“When more complexity is warranted, workflows offer predictability and consistency for well-defined tasks,” … “whereas agents are the better option when flexibility and model-driven decision-making are needed at scale.” ↑d
- 4Anthropic, Tool use with Claude, read at source 1 Oct 2026 (as cited on this site’s tool use meaning page): “Tool use (also called function calling) lets Claude call functions that you define or that Anthropic provides. Claude determines when to call a tool based on the user’s request and the tool’s description.”↑
- 5DeepLearning.AI, Agentic AI, course page, instructor Andrew Ng, undated, read at source 19 Oct 2026“You’ll master four design patterns that power agentic AI systems:” Reflection, Tool Use, Planning, Multi-Agent; “Reflection: AI critiques its own work and iterates to improve quality—like code review, but automated.”; “agentic workflows enable AI to plan multi-step processes”. ↑
- 6DeepLearning.AI, The Batch, issue 253, letter signed by Andrew Ng, 12 Jun 2024, read at source 12 Oct 2026 (as cited on this site’s agentic AI meaning page): “Unlike the noun ‘agent,’ the adjective ‘agentic’ allows us to contemplate such systems”; “we can acknowledge that there are different degrees to which systems can be agentic.”↑
- 7Kapoor, Stroebl, Siegel, Nadgir and Narayanan (Princeton University), AI Agents That Matter, arXiv, 1 Jul 2024, read at source 12 Oct 2026 (as cited on this site’s agentic AI meaning page): systems that “require less user supervision are more agentic”, as are systems “whose control flow is driven by an LLM, and hence dynamic”.↑
- 8Anthropic, Building effective agents, 19 Dec 2024, read at source 12 Oct 2026 (as cited on this site’s AI agent meaning page): “we recommend finding the simplest solution possible, and only increasing complexity when needed.”↑
- 9Anthropic, Building effective agents, 19 Dec 2024, read at source 12 Oct 2026 (as cited on this site’s AI agent meaning page): “The autonomous nature of agents means higher costs, and the potential for compounding errors.” “We recommend extensive testing in sandboxed environments, along with the appropriate guardrails.”↑
- 10OWASP, LLM01: Prompt Injection, GenAI / LLM Top 10, 2026 edition, read at source 11 Sep 2026 (as cited on this site’s AI agents for beginners page): prompt injection is LLM01, first on the list.↑
Read at source 19 Oct 2026: Anthropic, Building effective agents (19 Dec 2024); DeepLearning.AI, Agentic AI course page.
Reused from this site’s meaning pages and articles, with the read dates recorded there: IBM’s chatbot page (chatbot meaning, 1 Oct 2026); Anthropic’s tool use documentation (tool use meaning, 1 Oct 2026); The Batch issue 253 and the Princeton paper (agentic AI meaning, 12 Oct 2026); Anthropic’s simplest-solution and compounding-error lines (AI agent meaning, 12 Oct 2026); OWASP LLM01 (AI agents for beginners, 11 Sep 2026).
The comparisons on this page (friend on the phone, recipe card, assistant with keys), the table examples and the advice marked “this site’s reading” are this site’s own, not claims from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK