What Is Agentic AI? Andrew Ng's Course in 30 Minutes

TL;DR

An agentic AI workflow is any process where an LLM-based app executes multiple steps to complete a task, and Andrew Ng rejects the binary "is it an agent or not" debate in favor of a spectrum from less to more autonomous. The course covers five modules: agentic foundations, the reflection pattern, tool use, practical building tips including evaluations, and highly autonomous patterns. Wrapping a model in an agentic workflow beats calling it directly.

Transcript

I took Andrew Ning's agentic AI course for you. So, here's the cliffnotes version to save you the eight hours. I've been really looking forward to this one, but it is not enough for you just listen to me talk about stuff. So, at the end of this video, I have included a little assessment because research shows that immediately reviewing information ... Read More

Key Insights

• An agentic AI workflow is defined as a process where an LLM-based app executes multiple steps to complete a task, rather than a single direct model call. This definition is deliberately broad so that many multi-step LLM applications qualify as agentic.
• Andrew Ng explicitly rejects the binary notion that something either is or is not an agent, which is what people on X have debated for the past two years. His position is to call things agentic AI and place them on a spectrum from least to most autonomous.
• The spectrum of autonomy is a choice made by the person building the workflow. Less autonomous means predefined, clear steps and much more control over the final result. More autonomous means the agent makes more decisions itself and can produce creative results beyond what you imagined.
• The downside of high autonomy is less control and the possibility of results that are not what you are looking for at all. The trade-off is not that one end is better, but that control and creative range pull in opposite directions along the same spectrum.
• Wrapping a model in an agentic workflow performs much better than calling the model directly. It is also often faster and more modular, meaning components can be changed more easily, which is why agentic AI is popular and worth building.
• Andrew Ng's building blocks for agentic AI consist of just three components: models, tools, and evaluations. Models covers LLMs and other model types including multimodal, video generation, and audio generation models, so the framework is not LLM-only.
• Tools are the functions and capabilities given to an agent so it can perform different tasks. The categories named are external software access through APIs such as web search, real-time data and email checking, information retrieval from databases, and code execution for math, data analysis, and building.
• The practical way to design an agentic system is to ask how a human would accomplish the task, then translate each human step into either something the LLM can do natively or a tool it needs. In the customer email example, extraction and drafting are native while database lookup and sending require tools.
• Evaluations are the third component and are described as the other half of building an AI agent, because building an agent is not enough without verifying it works properly. Module four covers evaluations to a depth the presenter has not seen covered well anywhere else.
• An evaluation can be as simple as a small code snippet that checks a rule, for example compiling a list of competitor names such as Comp Co and Rival Co and checking whether any of them appear in the agent's drafted email response.

Questions & Answers

Q: What is agentic AI according to Andrew Ng's course?

Agentic AI is defined in the course as an agentic AI workflow: a process where an LLM-based app executes multiple steps to complete a task. Instead of calling a model once and taking its output, the application chains several steps together, such as planning, researching, drafting, reflecting, and revising. Andrew Ng deliberately avoids the binary question of whether something counts as an agent and instead treats any multi-step LLM-based process as agentic AI, positioned somewhere on a spectrum of autonomy.

Q: What are the five modules in the Agentic AI course?

Module one is an introduction to agentic workflows, covering the foundations of agentic AI that the rest of the course builds on. Module two is a deep dive into the reflection pattern, a common pattern for agentic workflows. Module three covers tool use, another common agentic pattern, along with many examples and ways to improve results. Module four gives practical tips for building agentic AI, including evaluations. Module five covers patterns for highly autonomous agents, more experimental workflows described as a taste for the future.

Q: What is the difference between a less autonomous and a more autonomous agentic workflow?

In a less autonomous workflow, you define the steps explicitly. For an essay on tea ceremonies, you direct the model to create an outline, decide whether web research is needed, do the research, write a first draft, identify parts needing revision, possibly request a human review, then produce the final draft. In a more autonomous workflow, you give the model the same goal plus a set of tools such as web search, news search, and archive search, and let it decide which tools to use, how to reflect, and whether to involve a human reviewer until it is satisfied.

Q: What are the building blocks of agentic AI?

Andrew Ng's framework uses three components. Models refers to the AI model itself, which can be a large language model or another type such as a multimodal, video generation, or audio generation model. Tools are the functions and capabilities given to the agent, including external software APIs, web search, real-time data, email checking, information retrieval from databases, and code execution for math, data analysis, and building things. The third component is evaluations, which check that the agent's output is actually correct.

Q: How do you design an agentic workflow for a real task?

Start by asking how a human would accomplish the task, then translate those human steps into what an LLM would do. For a customer email about a wrong item shipped, the steps are: extract key information such as the order number, the ordered blue blender, the received red toaster, and the weekend birthday timeline; find the relevant customer record; determine what went wrong; and send a reply. The LLM can extract information and draft the response by itself, but needs an orders database query tool and an email-sending tool for the other steps.

Q: Why use an agentic workflow instead of calling an LLM directly?

Wrapping a model in an agentic workflow makes it perform much better than calling it directly, which the video presents as a settled fact rather than a debate. Agentic workflows are also often faster and more modular, meaning individual components can be swapped or changed more easily without rebuilding everything. Those three benefits, better results, speed, and modularity, are given as the reason agentic AI is so popular and very much worth building.

Q: What are evaluations in agentic AI and why do they matter?

Evaluations are the third component of the framework and are about evaluating the results your agent produces. The video calls them the other half of building an AI agent: it is not good enough to just build an agent, you also have to make sure it works properly. Without evals, an agent handling a customer complaint might write a reply that boasts about being better than a rival company, or promise a refund without addressing the actual issue. Module four covers evaluations in depth.

Q: How do you write a simple evaluation for an AI agent?

One concrete example given is preventing an agent from mentioning competitors in customer emails, since that is in poor taste. You first compile a list of all competitor names you do not want mentioned, such as Comp Co and Rival Co, then write a small snippet of code that checks whether any competitor in that list appears in the agent's output. This is a rule-based, deterministic check rather than an LLM-based judgment, and it shows that evaluations do not have to be complicated to be useful.

Q: Which module of the Agentic AI course is the most important?

Module four, practical tips for building agentic AI, is described as by far the most important module. The reason given is its treatment of evaluations: the presenter says they have not seen this information, especially on evaluations, covered to this depth anywhere else. Evals are what everybody talks about but nobody has actually really covered well, and this course does it well. Module five, patterns for highly autonomous agents, is the more experimental content covering where agentic workflows are heading.

Summary & Key Takeaways

• The Agentic AI course has five modules: module one introduces agentic workflows as the foundation, module two is a deep dive into the reflection pattern, module three covers tool use with examples for improving results, module four gives practical tips for building agentic AI including evaluations, and module five covers more experimental patterns for highly autonomous agents.

• An agentic AI workflow is a process where an LLM-based app executes multiple steps to complete a task. An essay-writing example shows a less autonomous version with predefined steps: outline, decide whether web research is needed, research, write a first draft, identify parts needing revision, possibly request human review, then revise into a final draft.

• The more autonomous version of the same essay task gives the LLM a set of tools such as web search, news search, and archive search for research papers, then lets the model decide which tool to use, how to reflect on improvements, and whether to call a human reviewer until it is satisfied with the result.

• Andrew Ng's framework has three building blocks: models, tools, and evaluations. Models include LLMs plus other types such as video and audio generation models. Tools include external software APIs, web search, real-time data, email access, information retrieval from databases, and code execution for math, data analysis, and building things.

• A customer email example translates a human process into agentic steps: extract key information such as the order number and timeline, find the relevant customer record via an orders database query tool, determine what went wrong, then draft and send a reply using an email-sending tool. The LLM handles extraction and drafting natively.