YT YouTube2Text

YouTube2Text → Transcripts → AI Agents Fundamentals In 21 Minutes

AI Agents Fundamentals In 21 Minutes — Transcript

by Tina Huang · 4,432 words · 630 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00I learned about AI agents for you so
0:02here's the cliffnotes version to save
0:04you weeks of me learning about this
0:07there's not actually one course that
0:08just fully nicely covers everything so I
0:11did three courses wrote a bunch of
0:13papers and watch a lot of YouTube videos
0:15as well and of course actually made my
0:17own agents too my notes themselves are
0:19over 200 pages long but as per usual it
0:22is not enough just to listen to me talk
0:24about stuff so at the end of the video
0:25there is a little assessment which if
0:27you can answer these questions then
0:29congratulations you are now educated
0:31about AI agents now without further Ado
0:34let's get going a portion of this video
0:36is sponsored by HubSpot here's the
0:38outline first we're going to talk about
0:40what even are AI agents it is such a
0:42hyped up term now then we'll do a crash
0:44course on specifically multi-agent
0:47architectures it's really interesting
0:49developing field to make this actually
0:50all practical I'm going to then show you
0:52how to create an AI agent workflow which
0:55does not require any code I was honestly
0:57so shocked by how powerful and easy to
1:00use as well these workflows are then
1:01finally for those of you who are
1:02interested in getting into the field or
1:04even building your own AI agents for
1:06your businesses I will leave you with a
1:08piece of advice that when I heard it I
1:09was like holy so stay tuned for
1:12that at the end all right so let's first
1:15Define agents okay so believe it or not
1:17one of the most difficult things from
1:18this entire Deep dive into AI agents for
1:21me was just the actual definition of an
1:23AI agent probably because it's just such
1:25a new field and people are still trying
1:27to figure out what even it is and like
1:29how it works works so before watching
1:31this video if you were also confused I
1:33promise you it is not you let me walk
1:35you through this the easiest way to
1:36First Define ai agents is the given
1:38example of what is not an AI agent what
1:41is definitely not an AI agent is if you
1:43just ask an AI to do something for you
1:46otherwise known as one-hot prompting by
1:48the way if you're interested in leveling
1:49up your prompt engineering skills I did
1:51a video over here where I distilled down
1:53Google's 9-hour prompt engineering
1:55course into only 20 minutes so check it
1:57out anyways okay so what is definitely
1:59not an AI agent is if you're just asking
2:02AI to do something directly for example
2:04if you just go to chat gbt and write
2:06please write out an essay on topic X
2:07from start to finish in one go you'll
2:09still get a response and it'll still be
2:11like coherent and on topic but it'll
2:13probably also be quite vague and
2:16probably not what you were looking for
2:17on the other hand if you use an agentic
2:19workflow that will significantly improve
2:21your results and what that would look
2:23like is to break down that overarching
2:25task into different steps like first
2:27maybe writing an outline for the topic
2:29consider if you may need to do some web
2:30research then you might write your first
2:32draft consider what part of that draft
2:34may need more revision or more research
2:36revise your Draft before ultimately
2:38coming up with the essay a non- agentic
2:41workflow is just from start to finish
2:43and you're done while an agentic
2:44workflow is more a circular iterative
2:47process you think and you do research
2:49come up with an output and then you
2:50revise that and then you think and you
2:52do some more research come up with an
2:53output and you keep doing that until you
2:55get to your final result non agentic
2:57workflow straight up and down a gentic
2:59workflow
3:00circular okay so now let's add in a
3:02little bit of complexity you got your
3:04non- agentic workflow then you got your
3:05agentic workflow then you have a third
3:08level which is a truly autonomous AI
3:10agent this is when an AI can completely
3:12independently figure out the exact steps
3:15which tools to use go through that
3:16circular process of revising things by
3:18itself to finally come up with an output
3:21this is the level that we want our AI
3:22agents to become but currently as of the
3:25time of this filming at least we are not
3:27quite there yet we're still focusing on
3:30this second level of agentic workflows
3:32where there's certain agentic components
3:34to it but it's not fully autonomous yet
3:36but honestly with speeda AI is
3:38developing who knows maybe in like 2
3:39months that's going to happen we'll see
3:41Jarvis you there that's your
3:44Serv according to anging who's kind of
3:46like the Superstar of the AI World there
3:48are four massivly accepted agentic
3:50design patterns the first and simplest
3:52pattern is called reflection where
3:53you're simply asking an AI to more
3:56carefully look through its own results
3:58for example you might ask an AI to
3:59please write the code in order to
4:02complete you know a specific task and
4:04the AI is going to Output some code but
4:05you're not going to stop there you're
4:07going to ask the AI to please now check
4:09the code carefully for correctness style
4:11and efficiency and give constructive
4:13criticism for how to improve it the AI
4:15could look over its own code and then
4:16maybe find out that it made it a mistake
4:19on line five and in which case they can
4:21actually fix that line of code and
4:22continue improving its own output you're
4:24sort of helping that ai go through that
4:26circular agentic process to improve its
4:29output a very simple extension of this
4:31is instead of you being the one to help
4:34the AI figure this out you can actually
4:36create another Ai and have the other AI
4:39prompt the original AI to go through its
4:41own code and go through the reflection
4:43process so this is called a multi-agent
4:45framework and that's something that we
4:47will talk about a little bit later in
4:48the video and it's like a really really
4:50interesting field next up is tool use by
4:52giving an AI the ability to use tools
4:55you can help the AI better break down
4:57task and execute specific parts of the
4:59task for example if you're interested in
5:01buying a new coffee machine you can ask
5:03Nai what is the best coffee maker
5:05according to reviewers now if you give
5:07your AI the ability to search the
5:09internet like a web search tool you're
5:11allowing it to add in the steps of
5:13actually searching different reviews on
5:14the internet compiling them together
5:17before summarizing its findings which
5:19you would get a much better result than
5:20if you just ask it to directly come up
5:22with an answer another powerful commonly
5:24used tool is the code execution tool
5:27this allows your AI to actually create
5:29and to build build things like build out
5:30a website or calculate things things
5:32that involve numbers and math for
5:34example you can ask the AI if I invest
5:36$100 at compound 7% interest for 12
5:39years what do I have at the end your AI
5:42then can use this code execution tool to
5:44come up with the answer for you there
5:45are lots and lots of different tools
5:47that you can equip your AI with
5:49including object detection web
5:51generation ability to access your emails
5:53and your calendars to schedule events
5:55for you tool use is a very powerful
5:57agentic design pattern next up is
5:59planning and reasoning this is when you
6:01can give an AI a certain task that you
6:04want done and it's able to figure out
6:06what are the exact steps to accomplish
6:09these and what are the necessary tools
6:11that it needs in order to accomplish
6:13these steps for example you can ask an
6:15AI please generate an image where a girl
6:17is reading a book and her pose is the
6:19same as the boy in the image example.
6:21JPEG then please describe the new image
6:23with your voice with this agentic
6:25framework it's able to First Look at the
6:27image access a specific model to
6:29determine the pose of the boy in the
6:31image use another model to convert that
6:33specific pose to an image of a girl and
6:35another model to translate the image to
6:38text and finally a text to speech model
6:40to describe in audio what it is that the
6:43girl is doing a girl is sitting on a bed
6:46reading a book now finally we have
6:48multi-agent systems this is when instead
6:50of just having a single large language
6:52model a single AI do a certain thing you
6:55actually want to prompt different large
6:57language models to have different rules
7:00so the question you might have is like
7:01why can't you just have one Ai and just
7:03tell it to do everything right and the
7:05reason for this is that AI in this sense
7:07is actually quite similar to humans just
7:08like if you're trying to complete a
7:10project it's better to have a team of
7:12humans that all have their own
7:13specialized rules to come together to
7:15complete the project as opposed to just
7:17have like one person trying to juggle
7:19and handle everything same thing for AI
7:21there's research that shows by having
7:22this multi-agent workflow the results of
7:25the final product is generally better
7:26than just asking one AI to do all of it
7:29okay so here's a pneumonic in case you
7:30can't remember what the four agentic
7:32design patterns are just think about red
7:34turtles paint murals reflection tool use
7:37planning and multi-agents hint this will
7:41help in the little assessment at the end
7:42of this video okay so to make this all a
7:44little bit more concrete anding also
7:46showed us some tasks like some really
7:48cool tasks that were able to be
7:49accomplished by using these agentic
7:51design patterns for example like with
7:53this tool that has a agentic workflow
7:55built into it you can take an image of
7:57this soccer game and be able to identify
7:59Y and count number of players on the
8:01field you can also do stuff with video
8:03by prompting it given a video split the
8:05video into clips of 5 Seconds and find a
8:08clip where the goal is being scored
8:09display the frames associated with the
8:11goal that is pretty cool just thinking
8:13about the use cases you can do with so
8:15much video and image data that is
8:17currently untapped some other examples
8:19of a gentic systems that have produced
8:21really good results include AI powered
8:23research assistants that's able to
8:25research specific topics AI writers that
8:27can then write down these topics coders
8:29who can create software and personal
8:31assistance which I will actually show
8:33you how to build one later in the video
8:35as we see today AI agents and agentic
8:37workflows just like any other AI tool
8:40has a large component of prompt
8:42engineering it just shows that prompt
8:44engineering really is one of the highest
8:45Roi skills that you can learn today so
8:47if you're interested in leveling up your
8:49prompting skills I highly recommend that
8:50you check out this free prompt
8:52engineering Quickstar guide that I made
8:53with HubSpot it includes a step-by-step
8:56guide for creating great prompts and
8:57also tips to get better results my
8:59favorite part is that for all the
9:01examples there's a flow from bad to good
9:03to Great prompts to show how you can
9:05improve a prompt if you're able to go
9:07through this process and create great
9:08prompts you would just become so much
9:10more productive and get so much more out
9:12of AI so if you're interested do check
9:14it out at this link over here also
9:15linked in description thank you so much
9:17Hobs spa for creating this free resource
9:18with me and for sponsoring this portion
9:20of the
9:23video next up I want to do a quick crash
9:25course on multi-agent design patterns
9:28specifically this is where the 's a lot
9:29of focus and really cool breakthroughs
9:31that are happening I did a couple
9:33courses the best course that I found
9:34specifically for this topic was one by
9:36crew AI in collaboration with deep
9:38learning AI this course by crew AI gives
9:40a really good introduction to different
9:42types of multi-agent design patterns
9:45which I'm going to Now cover the first
9:47building block is a single AI agent and
9:49a single AI agent has four components it
9:52needs to have a specific task and answer
9:55what it's supposed to give you the model
9:56itself and tools that it has access to a
9:59nice little pneumonic here is tired
10:01alpaca's mix te task answers models
10:04tools for example you can have a travel
10:06planner AI agent its task is to plan a
10:093-day trip to Tokyo on a budget the
10:12answer that you want is a detailed itery
10:14with locations and cost as well as hotel
10:17bookings and any tickets the AI model
10:19could be anthropic CLA for example
10:21although you can switch that out for any
10:23other models that you like as well the
10:25tools that it needs include Google Maps
10:27Skyscanner for figuring out what the ti
10:29tickets are how much they cost
10:30booking.com for Logistics and your saved
10:33credit card informations so that you can
10:35actually place these bookings task
10:37answer model tools tired alpaca's mix te
10:40okay so we have our first singular unit
10:42of an agent and the simplest multi- aai
10:45agent would just be have two AI agents
10:47that work together on something each AI
10:49agent has its own programming but
10:51they're working together towards
10:53something an example of this would be a
10:55writer agent who is meant to write a
10:56blog article and an editor agent who is
10:59providing feedback for the writer even
11:02say with just two agents there's a
11:03couple interesting points here an agent
11:05can have its own task but an agent can
11:07also be working with another agent on a
11:09task while having its own task as well
11:11so there could be a lot of crisscross
11:13that's happening and for tools agents
11:15can have their own separate tools but a
11:18task can also have a tool which is
11:21really interesting you can actually
11:22program a task to have a specific tool
11:25so that an agent can only have access to
11:27it for that task and if you have more
11:30than one agent then you have a crew
11:32hence the name crew AI now when you add
11:35in additional agents there is even more
11:37complexity and it becomes really really
11:39interesting on how agents are
11:41interacting with each other I can go on
11:43for ages about all the different
11:44configurations of Agents working
11:46together and the tools that they're
11:48using but this course does give us a
11:50really nice kind of overview of the
11:52different design patterns that people
11:53have used and seem to be really helpful
11:55the first one is the sequential pattern
11:57this is the simplest when you just have
11:59one One agent do something and then it
12:00passes it on to another agent that does
12:02something else and another agent that
12:03does something else sort of like an
12:05assembly line an example it has would be
12:06AI powered document processing you can
12:09have your first agent which extracts
12:11text from scan documents that it passes
12:13on to another agent who summarizes the
12:15text then passes on to the next agent
12:18who then extracts action items and puts
12:19it into a summary and finally to a
12:21fourth agent that saves the data into a
12:24database a higher article higher AR a
12:28higher AR h two hours later higher
12:31article agent system would have a leader
12:34or manager agent that supervised
12:36multiple agents that have their own
12:38specific task these sub agents will
12:40complete their task and Report their
12:41results back to the manager agent who
12:43then compiles it all together an example
12:45of this would be writing a report for
12:47business decision-making you have your
12:48manager AI agent that receives this task
12:51and then delegates it to different sub
12:53agents sub agent one monitors and
12:55reports back market trends and it would
12:57have specialized tools for looking into
12:59these markets sub agent 2 could be
13:01monitoring internal customer sentiment
13:04so has access to the internal databases
13:06to see what kind of feedback customers
13:08are giving while sub agent 3 tracks
13:10internal metrics across the company so
13:13it's understanding how this specific
13:15product is interplaying with other
13:17products within the company now after
13:19all these agents do their job they would
13:20all report back to the manager agent
13:22who's able to combine everything
13:24together and it might actually pass this
13:26along to another agent say like a
13:28decision making agent who may aggregate
13:30different insights and professionally
13:32put it into a report and come up with a
13:34ultimate business decision next up is
13:36the hybrid system this combines
13:38different sequential and hierarchical
13:41structures together agents can
13:42collaborate top down as well as
13:44sequentially an example of this would be
13:46in autonomous vehicles at the top level
13:48you might have a AI agent that plans the
13:50overall route and traffic strategy for
13:52an autonomous vehicle then you have the
13:54sub agents that handle things like
13:56real-time Sensor Fusion collision
13:58avoidance
13:59and road condition analysis but it's not
14:01enough just to aggregate this
14:03information together and then just give
14:04it to the top level AI because you need
14:06to have a continuous feedback loop as
14:08the vehicle itself is moving and the
14:10road conditions and everything around it
14:12internally and externally is all
14:13changing as well you need to have lots
14:15of different little feedback loops
14:17between these different agents and then
14:18communicating continuously with the top
14:21level agent as well this design pattern
14:22is really common in things like robotics
14:25navigation systems and adaptive AI
14:27systems basically like in places where
14:29there's lots of moving Parts there are
14:31also parallel agent Design Systems this
14:33is when you have agents working on
14:35different work streams independently
14:36agents would be handling different parts
14:38of a task simultaneously often to speed
14:40up processing an example of this would
14:42be like AI for large scale data analysis
14:45this is a very common structure the very
14:47large analysis involves different
14:49components and agents will take chunks
14:51of that data and process them separately
14:53ultimately at the end merging everything
14:55together and finally there's
14:56asynchronous multi-agent systems this is
14:59when agents execute tax independently
15:01and at different times this is a system
15:03that's proven to handle uncertain
15:05conditions better than sequential or
15:07parallel approaches an example of this
15:09would be something like an AI powered
15:11cyber security threat detection you got
15:13agent one that's monitoring Network
15:15traffic in real time agent two that's
15:17monitoring suspicious usage patterns and
15:19agent three that's just randomly
15:20sampling and testing out different use
15:22cases when any of these agents picked up
15:24something anomalous they would flag it
15:26and then other things would happen after
15:27that this type of AC synchronous design
15:29pattern is especially helpful for
15:31anything that requires real-time
15:33monitoring or self-healing systems and
15:35finally to put them all together you can
15:37actually have these different systems
15:39and then link up these systems
15:41themselves and this is called a float
15:43this can result in really complex and
15:45interesting processing and results but
15:47the note to make here is that as you
15:50increase the complexity of these systems
15:52you're also basically increasing the
15:54amount of chaos that's within it as well
15:56since you don't actually have like
15:58Direct access to these agents right like
16:01you can provide them with feedback and
16:02there's ways of doing that but as you
16:04add on more and more complexity there's
16:06more things and more moving parts that
16:08are kind of just like interacting with
16:09each other it's actually pretty similar
16:11to how human companies work right the
16:13bigger your company becomes the more
16:15chaotic it starts becoming as well and
16:17the more emphasis you need to place on
16:18like hierarchies and different you know
16:20organization structures I don't know
16:22this for sure but if I were to bet I do
16:25think a lot of research that people do
16:27into systems like human systems and
16:29companies probably also comes into play
16:31for multi-agent AI systems too for the
16:34rest of the course they basically go
16:35through different implementations and
16:36examples for these different multi- aai
16:38agent systems so instead of going
16:39through all of these examples I'm just
16:41going to link in the description some of
16:42these notebooks where you can use code
16:44to implement these systems using crew AI
16:47but do not worry if you're not a coder
16:48where you're just not interested in
16:49coding I'm actually going to now show
16:51you a way of creating these multi- aai
16:54agent systems completely with a no code
16:56tool called n8n robot building sequence
16:59activated I'm so glad we tried out our
17:01new Android building device instead of
17:03using that old dinosaur some of you guys
17:05may have heard of make.com which people
17:08also use to make these multi- aai agent
17:10systems um but na an is actually better
17:12for doing this specifically credit here
17:14to David Andre's 40-minute tutorial
17:16which is what I follow and adapted to
17:18create my own AI assistant this is a
17:20telegram based AI assistant that's able
17:23to communicate with you and help you
17:25prioritize your task by accessing your
17:26Google calendars and it can also create
17:28calendar events for you so you can go on
17:30Telegram and talk to Inky bot which is
17:32the assistant's name and say what do I
17:34need to do today and it tells me that
17:36today is February 5th 2025 and I have to
17:39film this video and the time is from
17:4212:00 p.m. until 400 p.m. in Hong Kong
17:44and it also asked me to list what are my
17:46other priorities for today so that it
17:48can come up with a list of tasks and
17:50prioritize it for me so I'm just telling
17:53that filming is my greatest priority and
17:54have these other things so it's able to
17:56prioritize and put in sequence my other
17:58tasks as as well as actually schedule
17:59calendar events corresponding to these
18:02specific task okay so the way that this
18:04flow works is first you have the
18:06telegram trigger so this is when I send
18:08a message to Inky bot and from there
18:10there's a switch um this is because it
18:12can take both text and voice input so if
18:15it's text input you would just directly
18:17take that information and feed it into
18:18the AI agent but if it's voice input we
18:21first get telegram to get the file send
18:23it to open AI to transcribe the file and
18:26then send the text information to the AI
18:28agent as well now the AI agent here is
18:30the interesting part remember tired
18:32alpacas make tea the task is taking the
18:35user's query asking about what needs to
18:37be done for today the answer is a
18:39prioritized to-do list as well as
18:42scheduled events into Google Calendar if
18:44needed the model we're using here is
18:45open AI GPT 40 mini but you can also
18:48change that out for whatever other model
18:50that you want as well like Claud Gemini
18:52llama deep seek whatever you like and
18:54finally it has two different tools the
18:56first tool is the get calendar events so
18:58it's able to read the Google calendar
19:00and see what events there are for the
19:02day it can also create calendar events
19:05so when the user wants to add other
19:06events into the list it can then go and
19:09actually create these events on the
19:11Google Calendar yeah and then it would
19:13be able to communicate through telegram
19:15with the user until it comes up with a
19:18list that the user is happy about they
19:20can also do things like check off the
19:21list plan ahead look at what happened in
19:24the past a lot of other things as well
19:26as you can see just the single agent the
19:27super simple work flow can already
19:29produce really cool results so think
19:31about adding other agents there other
19:34functionalities it's really really cool
19:36what you can do with this and it's
19:38totally no code which is
19:45crazy all right final section is on the
19:47opportunities for AI agents I watched a
19:50lot of YouTube videos and read a lot of
19:52Articles mostly for this section and the
19:55biggest takeaway that I got from this
19:56like assuming you want to be building
19:58something thing using AI agents
20:00something that is useful for other
20:01people you're building up a business is
20:03from this why combinator video where
20:05they say that for every SAS or software
20:07as a service company there will be a
20:10corresponding AI agent company let me
20:12just like repeat that because this is
20:14like huge guidance in terms of what to
20:15build for every software as a service
20:18company like all the software service
20:19companies that we see today there will
20:21be a corresponding AI agent version of
20:23that so if you don't know what to build
20:26or what to do right now and you want to
20:27play around with a agents just literally
20:30take a SAS company and then think about
20:32how do I make that into an AI agent
20:34company just ask chachu BT what are some
20:36top SAS companies says Adobe Microsoft
20:39Salesforce Shopify link tree canva
20:43Squarespace and on and on and on and on
20:45there are so many literally every
20:47company that is a sass unicorn you could
20:49imagine there's a vertical AI unicorn
20:52equivalent I really think that piece of
20:54advice is literal gold let me know in
20:56the comments if there's a specific AI
20:58agent that you're interested in building
20:59or an AI agent business all right we
21:02have come to the end of this video thank
21:04you so much for watching through it as
21:05promised here is a little assessment if
21:07you can answer all these questions then
21:10congratulations you can consider
21:11yourself educated on AI agents let me
21:14know in the comments what other topics
21:15whether that's like AI topics or other
21:17topics is fine as well that you want me
21:19to do a deep dive into all right thank
21:22you all so much for watching and I will
21:23see you guys in the next video where
21:25live stream

About this transcript

This page contains the full transcript of AI Agents Fundamentals In 21 Minutes by Tina Huang, generated from the public captions YouTube serves with the video. The transcript has 4,432 words across 630 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.