An agentic workflow is a way of using AI in several steps instead of one: the model plans, looks things up or uses tools, checks its work and revises it before you get the result. The steps may be laid out by a person or chosen by the AI as it goes.
Think of the difference between asking a friend to write a speech on the spot and asking them to draft it, check the facts, read it back and fix the weak parts. Same friend, much better speech. A plain chatbot answer is the speech on the spot. An agentic workflow builds in the drafting, checking and fixing.
In a sentence: “Our report tool runs an agentic workflow: it outlines, searches our files for each section, drafts, then has a second pass check every figure against its source.”
How is an agentic workflow different from an AI agent?
The plain point: in a workflow a person sets the route; in an agent the AI picks the route itself. Both count as agentic.
| Term | Who decides the steps |
|---|---|
| Single prompt | Nobody: one question, one answer, no checking or revising. |
| Agentic workflow | Usually a person: the steps are set out in advance, and the AI does each one. |
| AI agent | The AI: it decides what to do next, which tools to use and when it is done. |
| Agentic AI | Umbrella word: any system that works towards a goal with limited supervision, workflows and agents alike. |
Anthropic, which makes Claude, draws this line. It puts both under “agentic systems”, then splits them: “Workflows are systems where LLMs and tools are orchestrated through predefined code paths.” Agents are systems where models “dynamically direct their own processes and tool usage”.1 People do use “agentic workflow” loosely for both, so when someone says it, ask who decides the steps. Researchers at Princeton describe agentic as “a spectrum” rather than a yes-or-no label.2
What did Andrew Ng mean by agentic workflows?
The plain point: let the model work over a task several times, the way a person drafts and redrafts, instead of answering in one go.
Andrew Ng wrote about agentic workflows in The Batch, the DeepLearning.AI newsletter, in March 2024. He compared normal use to writing an essay “typing straight through with no backspacing allowed”, while an agentic workflow lets the model “iterate over a document many times”: outline, research, draft, then revise.3 His view: “With AI, such an iterative workflow yields much better results than writing in a single pass.”4 He named four design patterns: reflection (the AI critiques its own work), tool use, planning, and multi-agent collaboration.4
What does an agentic workflow look like?
The plain point: a few repeatable shapes, from a simple chain of steps to one AI checking another.
Anthropic names five workflow patterns: prompt chaining, routing, parallelization, orchestrator-workers and evaluator-optimizer.1 In plain words:
- Prompt chaining: each step’s output feeds the next, such as outline, then draft, then edit.
- Routing: the first step sorts the request and sends it down the right path, such as billing questions one way and technical ones another.
- Parallelization: several parts run at once and the results are combined.
- Orchestrator-workers: one AI splits the job up and hands pieces to others.
- Evaluator-optimizer: in Anthropic’s words, “one LLM call generates a response while another provides evaluation and feedback”, in a loop.1
An agentic workflow trades speed for quality: more steps, more checking. Use the simplest setup that does the job, and add steps only when a single answer is not good enough.
Why it matters to you
The plain point: more steps can mean better work, but also more cost and more places for errors to creep in.
Anthropic recommends “finding the simplest solution possible, and only increasing complexity when needed”, and says that when more is needed, workflows “offer predictability and consistency for well-defined tasks”.1 It also warns that more autonomy brings “higher costs, and the potential for compounding errors”.5 In this site’s reading, a mistake in step two is carried into every step after it, which is why a check step, and a person reviewing the final result, are worth building in.
Questions people ask
Is an agentic workflow the same as an AI agent?
Not quite. In Anthropic’s terms a workflow follows “predefined code paths”, while an agent directs its own steps; both are agentic systems.1
Can I run an agentic workflow in an ordinary chatbot?
A simple one, yes, by hand. In this site’s use: ask for an outline, then a draft, then ask the chatbot to list the draft’s weaknesses, then ask it to fix them. That is Ng’s draft-and-revise idea done one message at a time.
Are agentic workflows always better than one prompt?
No. They are slower and cost more to run, and Anthropic advises the simplest setup that works, adding complexity only when needed.1
Where to go next
- The umbrella word: what is agentic AI?
- When the AI picks the steps: what is an AI agent?
- The simplest pattern: what is prompt chaining?
- How the AI acts at each step: what is function calling?
- Getting started: AI agents for beginners and loops that improve and loops that degrade.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What makes a workflow agentic rather than a single prompt?
- It uses a bigger model
- The AI works in several steps, such as planning, checking and revising
- It runs without a person ever seeing the result
Show the answer
Answer: The AI works in several steps, such as planning, checking and revising. An agentic workflow builds in steps like drafting, checking and fixing instead of one answer.
2. In Anthropic’s terms, how does a workflow differ from an agent?
- A workflow follows steps set in advance; an agent decides its own steps
- A workflow has no AI in it
- An agent always needs a person to set each step
Show the answer
Answer: A workflow follows steps set in advance; an agent decides its own steps. Workflows follow predefined paths; agents direct their own process and tool use.
3. What is a downside of adding more steps?
- The AI forgets how to write
- Nothing; more steps are always better
- More cost, and an early mistake can carry into every later step
Show the answer
Answer: More cost, and an early mistake can carry into every later step. More steps can improve quality but add cost and room for errors to compound.
Related terms: Agentic AI · API · Benchmark · Coding agent · Vector database · Vibe coding
- 1Anthropic, Building effective agents, 19 Dec 2024, read at source 19 Oct 2026“we categorize all these variations as agentic systems”; “Workflows are systems where LLMs and tools are orchestrated through predefined code paths.”; “Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage” (sentence continues). ↑adescribes the workflows prompt chaining, routing, parallelization, orchestrator-workers and evaluator-optimizer. ↑b“In the evaluator-optimizer workflow, one LLM call generates a response while another provides evaluation and feedback” (sentence continues). ↑c“we recommend finding the simplest solution possible, and only increasing complexity when needed”; “When more complexity is warranted, workflows offer predictability and consistency for well-defined tasks” (sentence continues). ↑d“Workflows are systems where LLMs and tools are orchestrated through predefined code paths.” ↑e“we recommend finding the simplest solution possible, and only increasing complexity when needed.” ↑f
- 2Kapoor, Stroebl, Siegel, Nadgir and Narayanan (Princeton University), AI Agents That Matter, arXiv, 1 Jul 2024, read at source 12 Oct 2026 (as cited on agentic AI meaning): “Many of them view it as a spectrum — sometimes denoted by the term ‘agentic’”.↑
- 3DeepLearning.AI, The Batch, issue 241, letter by Andrew Ng, 20 Mar 2024, read at source 19 Oct 2026“we mostly use LLMs in zero-shot mode”; “typing straight through with no backspacing allowed”; an agent workflow lets the LLM “iterate over a document many times”; “Revise the draft taking into account any weaknesses spotted.” ↑
- 4DeepLearning.AI, The Batch, issue 241, 20 Mar 2024, read at source 19 Oct 2026“With AI, such an iterative workflow yields much better results than writing in a single pass.” ↑anames Reflection, Tool use, Planning and Multi-agent collaboration as design patterns. ↑b
- 5Anthropic, Building effective agents, 19 Dec 2024, read at source 12 Oct 2026 (as cited on agentic AI meaning): “The autonomous nature of agents means higher costs, and the potential for compounding errors.”↑
DeepLearning.AI, The Batch, issue 241, letter by Andrew Ng, 20 Mar 2024, read at source 19 Oct 2026.
Kapoor, Stroebl, Siegel, Nadgir and Narayanan (Princeton University), AI Agents That Matter, arXiv, 1 Jul 2024, read at source 12 Oct 2026 (as cited on agentic AI meaning).
The comparisons and examples on this page (the speech, the report tool, the who-decides table, the chatbot method) are this site’s own ways of explaining, 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