What Is Agentic AI? From One Prompt to a Workflow
My notes from Andrew Ng's Agentic AI course on DeepLearning.AI, and the first agent I designed from them.
The short answer
Agentic AI means you stop asking a language model for the finished result in one go. You split the job into steps, let the model and a few tools handle one step at a time, and check the result along the way. This loop of planning, acting, checking and improving gives better results than a single prompt, even with the same model.
Why one prompt is often not enough
Most people use an LLM in one move: write a prompt, read the answer, copy it. That works for a short email. It works less well for a research report or a block of code, because you are asking for the whole thing in one pass, like typing a long document without ever using the backspace key.
People don't work like that. Take a cover letter. You read the job ad, note what matters, write a draft, read it again, cut the weak parts, and maybe ask a friend to check it. The final version is better because it went through steps.
What an agentic workflow does
An agentic workflow copies that way of working. A typical one has five parts.
- 1
Plan
Break the task into steps or write an outline.
- 2
Use tools
Search the web, fetch a page, query a database or run code.
- 3
Draft
Give what the tools found to the LLM and let it write a first version.
- 4
Reflect
Ask the LLM, or a second LLM, to find gaps and errors in the draft, then revise.
- 5
Check with a human
A person approves important results before they go out.
This takes longer than one prompt, and it costs more API calls. In return you get better quality, and you can look at every step when something goes wrong.
Does it really work?
Andrew Ng gave a coding example in The Batch in March 2024. On HumanEval, a set of programming problems, GPT-3.5 with a single prompt solved 48.1%. GPT-4 with a single prompt solved 67.0%. GPT-3.5 inside an agent workflow reached up to 95.1%. The results came from several research teams, and Ng's article summarises them. An older, smaller model with a good workflow beat a stronger model used alone.
These are 2024 numbers on one benchmark, and newer models have moved on since. Treat them as a lesson in workflow design. They say nothing about which model is best today.
Workflow or agent?
Two terms get mixed up here. Anthropic separated them in its December 2024 guide "Building effective agents". Both count as agentic systems. In a workflow, your code decides the order of the steps and the LLM does the work inside each step. In an agent, the LLM decides the steps and chooses the tools.
Ng prefers to see this as a scale rather than a yes-or-no question:
- Less autonomous: the developer fixes the steps and the tools, and the LLM only writes text.
- Semi-autonomous: the tools are fixed, and the LLM picks which one to call.
- Highly autonomous: the LLM plans the steps, chooses tools, and can even write new ones.
Anthropic's advice is to start with the simplest solution that works and add complexity only when it is needed. Workflows suit well-defined tasks where you want predictable results. The next lesson in this track goes deeper into the scale.
The four design patterns
Ng groups the building blocks of agentic systems into four patterns: reflection, tool use, planning and multi-agent collaboration. They get a lesson of their own in this track.
My first example: the Job-Match Agent
I'm looking for AI Engineer or AI Automation roles in the DACH region, so my first agent is a job matcher. I have drawn the workflow. I haven't built it yet.
The Job-Match Agent workflow in order:
- Input: the CV.
- LLM: extract skills, keywords and job titles from the CV.
- LLM: write 3 to 5 job search queries.
- Web search tool: find job links.
- Web fetch tool: get the full job descriptions.
- LLM: compare the CV skills with each job and give a fit score from 0 to 100, with reasons and skill gaps.
- Code filter: keep only jobs with a score of 70 or higher.
- Output: matching jobs as JSON with score, reasons and skill gaps.
It starts with my CV and runs through six steps:
- LLM: extract skills, keywords and job titles from the CV.
- LLM: write 3 to 5 job search queries.
- Web search (tool): find job links.
- Web fetch (tool): get the full job descriptions.
- LLM: compare the CV skills with each job and give a fit score from 0 to 100, with reasons and skill gaps.
- Code filter: keep only jobs with a score of 70 or higher.
The output is a JSON list of matching jobs with score, reasons and skill gaps.
Three of the six steps are LLM calls, and three are tools or plain code. The order is fixed in code and the LLM only reads and writes text, so this agent sits at the less autonomous end of the scale. The last step is plain code because "70 or higher" is a rule, and code applies a rule the same way every time.
Some questions I can't answer yet. Is 70 the right cut-off? Will the model give the same job the same score on two runs? Are the extracted skills correct? These are evaluation questions, and they get their own lesson.
What comes next in this track
The spectrum of autonomy, the four design patterns, task decomposition (how to get from "find me a job" to six steps like the ones above), and evaluation. After that I'll build the Job-Match Agent step by step and write down what goes wrong.
Test yourself
Take a moment to guess each answer in your head, then open the correct answer and compare.
Question 1 of 5
What is agentic AI in simple terms?
Show the correct answer
Correct answer. Agentic AI is a way of using language models where the work is split into several steps, with tools and checks in between, instead of one prompt and one answer. The model can plan, search, draft and review its own work before a result reaches you.
Question 2 of 5
What is the difference between an agentic workflow and an AI agent?
Show the correct answer
Correct answer. In an agentic workflow, the developer's code sets the order of the steps. In an AI agent, the LLM decides the order and chooses the tools itself. Anthropic calls both agentic systems and recommends starting with the simpler, more predictable workflow.
Question 3 of 5
Is an agentic workflow always better than a single prompt?
Show the correct answer
Correct answer. No. The quality is usually higher, but it takes longer and costs more. In the HumanEval example, GPT-3.5 in an agent workflow reached up to 95.1%, against 67.0% for GPT-4 with one prompt. Anthropic also notes that for many applications, one well-built LLM call with retrieval and examples is enough.
Question 4 of 5
Do you need a framework to build an agentic workflow?
Show the correct answer
Correct answer. No. Andrew Ng's Agentic AI course on DeepLearning.AI builds each pattern in plain Python first, then looks at frameworks. Plain code gives you full control over the steps, the tools and the evaluation.
Question 5 of 5
Look at the Job-Match Agent. Which of the six steps is plain code, and why?
Show the correct answer
Correct answer. Step 6, the code filter. "70 or higher" is a rule, and code applies a rule the same way every time. Steps 1, 2 and 5 are LLM calls, and steps 3 and 4 are tools that search and fetch.
Sources
- Andrew Ng, Agentic Design Patterns Part 1 (opens in a new tab), The Batch, DeepLearning.AI, 20 March 2024.
- Anthropic, Building effective agents (opens in a new tab), 19 December 2024.
- DeepLearning.AI, Agentic AI course by Andrew Ng (opens in a new tab).
Updated 17 October 2026
Update log
- : Added a link to lesson 2, which goes deeper into the scale of autonomy.
- : First published.
Follow the build log
Comments
No comments yet — be the first to add one.