Unknown

Hello friends how are you guys doing can I get a mic check you good yeah okay cool how are you guys doing it's been a while this is Hong Kong I want to show you guys so this is the place I I'm living in Hong Kong now there's a mountain over there yeah I have moved to Hong Kong hello greetings from Pennsylvania hello hi from Spain hello there we go have to refresh to get it to load cool all right well I'm going to actually get started pretty quickly because there's quite a lot of stuff

to get through so we're doing a crash course today um I'm trying my best to get through most of it but if we it we don't get to the demo portion and how to actually build it we can do like a follow-up session as well and happy New Year Happy Chinese New Year for anybody that does Chinese New Year um all right let me see okay so today um I'm GNA be doing the a crash course oh shoot how do I I always have this problem if I do present interview then I cannot see my

own then I cannot see my own things so I guess I cannot do present interview has to be a better way of doing this honestly does that work then it gets kind of weird right okay I'll try my best see if this works will you leave this live up your voice is a bit low gain I always have that is it better now the gain is maximum now okay all right let's get started so what I'm going to talk about today is as many of you probably know AI agents like a lot of people are

talking about AI agents right now and you know they're like oh you should make AI businesses you should make AI agents for your productivity you should do this you should do that all these different things are being developed blah blah blah you know there's always like all of this going going on um but AI agents is one of those things after like I did a significant amount of digging it's kind of hard sometimes to distinguish between what is actually like a real fundamental important AI thing as opposed to whatever the next hype object is right

and AI agents is absolutely a movement that is trending in a way that it actually makes sense like it's being integrated into businesses there's actually a lot of business opportunities for AI agents so it absolutely is something that is worth learning as opposed to some other like you know whatever hype news of the new tool and things are are about so the thing is like with AI agents because it's so new there aren't really like that many good resources in terms that that fundamentally covers everything right so I actually did my own kind of digging

I like read a lot of books um a lot of Articles made my own agents U I've been playing around with it like just AI in general for for quite a while but actually like creating the agents and stuff I've been doing that so this is kind of like the crash course version of everything you need to know and I could just show you a demo and be like oh like this is how you can build an AI agent right but I actually want to go a little bit more fundamental than that I actually want

to explain to you um from a lot of different resources like how what it is AI agents are like what the actual fundamental Trends are what does it mean where things are going and then if we have time but maybe next time I want to show you how to actually build an AI agent does that make sense like I don't want to just like show you guys this is how you build an AI agent I actually really want you to understand what it is and the potential that it has has and what the components are

in a more like systematic way okay so as per usual you know do whatever you just message I'll be like monitoring to chats and things like that all the time and yeah like if you have any questions like feel free to interrupt me or every Point um also while we're at it so I have it here I'll put it here again if you could if you want the PowerPoint slides afterwards please like click that link and you can sign up for the workshop email so we're all we're not going to spam you just email you

like yeah we won't spam you it's just going to be once a week where I'll email I'll email you the slides and also the Google Calendar invites for future workshops as well yeah and it's completely free so yep all right let us get started AI agents crash course let me see any comments uh can we still see yep yep this is going to be up later as well so do not worry it's all going to be up still all right okay AI agents crash course so here's the agenda we're just going to talk about what

is an AI agent surprisingly it is quite difficult to Define just because like yeah again it's like because it's so new and people have a lot of different approaches to it so I'm going to try to Define it for you guys um I'm going to talk about different types of agentic workflows um and agentic design patterns and then specifically I want to talk about multi-agent design patterns this is kind of like the really cool thing these days it's how agents are working with each other in order to create different pieces of software uh in order

to do different types of tasks so this is kind of like where a lot of excitement is being focused on and there's a lot of potential um in in this specific thing and then I'm going to I'm doing a practical implementation like show you guys um what it actually looks like and hopefully be able to go through that if not we can do that next time as well and then I also want to hopefully we have time go through some of the business opportunities um using AI agents as well but potentially that could be like

a part part two for practical the part three of your business opportunities but the thing that I definitely want to cover uh during this session this Workshop is going to be like the things over here the fundamental um information okay so first let's actually understand what AI agents are and really like I've seen like so many definitions of AI agents let me like make my face go somewhere else yeah there's like so many definitions of AI agents but the easiest way is actually to talk about like what is not an AI agent right so if

you're just like asking AI whether chat to people CL whatever you're just asking AI hey like write an essay for me in one go this is called ones shot prompting and this is definitely not an AI agent you're just simply asking AI in order to do something for you in one go right and you can get like a decent amount of like a good a decent result from this but it's not very good it's probably going to be quite vague um so in contrast to this you have something called an agentic workflow and you can

think about an agentic workflow as more of an iterative process think about like circular process where you're asking an AI to break down different task and then revise outputs that improve different results so the way I like to think about it is think about um an a not an AI agent is something that's like up and down right straight it's like do this for me write me an write me an essay while an agentic workflow is circular so you're saying can you write the outline for this essay for me can you uh revise the outline

do some more research write the actual essay correct essay and keep going in circles so that is the defining feature of an agentic workflow you're providing it feedback and you're giving it information in order to iterate and to get better results from that okay and then this is kind of where we're at right now in terms of agentic workflows the ideal goal that we're trying to go for is what is called fully autonomous agents it's not something that we have right now but this would when you're able to just say like give an AI like

a specific task and is able to do it completely by itself like for example uh it would be like you know stuff in the movies like Jarvis or something like that just like tell to do something and then it's able to completely execute it if you had a marketing AI just go like make me a marketing campaign and then it's able to like go through that entire process using tools using steps to create like an optimal solution to this that's like where we want to be headed um but currently we are not there we're still

in the process of a gentic workflows kind of like what happened is that first you know when we started playing around with AI uh we started just doing like you know one shot prompting prompt engineering just like that and then we realize oh like if we have agentic workflows and it's a circular thing maybe with multiple agents working with each other it's going to get better results and then at some point we're going to get fully autonomous agents where you're able to just give it a task and it's able to figure out all the steps

by itself use the different tools that it needs to do and then ultimately get to a final result okay so yeah this is like a diagram in terms of how you can think about it um oops sorry so a non- agentic workflow is going to be one p execution no adjustments and examples writing an essay to start to finish a gentic workflow is going to be like uh giving a stepbystep execution there's a feedback loop an example would be writing an outline doing research Drafting and revising and a truly autonomous agent over here is going

to be like fully independent and do everything by itself I'm going to stop here and see if anybody has any questions is it also about agents yes we're talking about agents right now um theama course by free code I'm not sure about linking in her class I am not sure what we're talking about here uh can we see this later grade is it after yep it is still going to be live and yes it's on the pin comments if you sign up for that we'll also send you the slides over cuz like yeah I I'll

send you the slides over uh the live stream itself will be here as well so do not worry um do you use open source models yeah I mean personally you can play around with a lot of different models um I actually think what's also really cool about agentic workflows uh it's it's designed to be model agnostic so you can actually plug it in agents are more like a framework if that makes sense it's more of a framework for how you do things and you can change to different models you can change to different tools and

you can change things around depending on what your task is supposed to be and what kind of agent it is that you are that's why I'm so keen on you know G explaining to you the fundamental Frameworks um of doing this as opposed to just you know purely showing you how to do it with a specific type of model um what's the difference between AI agents and gentic AI yeah great so AI agents so that's why I actually did the way I I wrote it the way it is it's like very hard to Define because

we're still trying to figure it out I think the easiest way of seeing it is by this three step process where what is definitely not an AI agent a gentic workflow means it has like certain agent capabilities right that iterative process or certain agentic agent likee iterative things and then an AI an actual AI agent like by definition is a fully autonomous AI agent that is going to be able to completely do something by itself so that's the distinguishment um and right now we're still at a stage of more agentic workflows like there's components of

Agents but we're not completely at an autonomous agent yet if you heard about like Devin when the like first autonomous AI coder came out they said that it was going to be like a fully autonomous thing but it's really not because we're still like not quite there yet but that's like the ultimate goal for what we want okay um when do you think fully autonomous agents will be available for the public to use I don't know you know thing is with AI it's so hard to predict it could be like a month it could be

like a few months it's I I yeah I was really hard to predict but I do think it is progressing significantly and I'll talk about this a little bit more in terms of like the actual Frameworks it's probably going to look something like multiple agents working together to come up with this like fully autonomous agent kind of look that's probably what's going to happen all right let's keep moving oops um okay so in case you cannot remember what are the components uh what are the agentic design patterns so I'm going to talk be talking about

these like agentic design patterns now um and this is courtesy of Andrew ning he's like a superar in AI so I'm not the one claiming credit for this but I did come up with this pneumonic which we can revise later so the four different um different design patterns like the agentic design patterns that are what people are looking into right now it's called reflection tool use planning and multi-agent Frameworks so I like to use neonics but can't remember stuff so I like to think of like red turtles paint murals so we actually review this so

to remember like R for reflection Turtles for Tool use paint for planning and murals for multi-agents so in the future you know I'll probably give you a little quiz at the end you can't remember what agentic design patterns are it's called red turtles paint murals reflection tool Ed planning and multi-agents right I'm going to go to each of these um now so a gentic design patter let first talk about reflection so reflection is the most simple form of a gentic design pattern and this is when AI reviews and improves its own output by analyzing and

refining a response so this can be like this is honestly super simple you might not even think about it as an agent because it's so simple um but for example if you're trying to write a code review for example you're writing you're asking hey can you like write me this piece of code and then you would say can you write the code for a specific task and then you can then ask the AI to actually check its own code to make sure that it is doing what it's supposed to be doing um and then identifies

its own errors and then you can tell that oh after you identified this error please make these improvements and then ultimately it would make that Improvement and correct its own process so this is the simplest version you're asking the AI to reflect on what it has done iterate on that and then um be able to come up with a better solution and you can kind of see like you know this is we'll talk about the multi-agent framework later but this is kind of like the simplest one and you can definitely see instead of you as

a person prompting it to go and do something right like instead of you prompting it to go review its code check errors you can easily see how it is that you can also ask another AI like another AI agent to be checking the code for that person and and asking the agent to reflect so you can have two different agents now one of them is writing the code and the other agent is asking the original AI agent in order to check the code does that make sense so reflection is the simplest approach uh let me

see if anybody has many comments this is all new info to me I always thought AI agents were the bad guys in The Matrix I think that would be like the fully autonomous AI agents if we ever get to that point point they will be the bad guys in The Matrix um I see the main difference between an agent and a rag is that the agent has access to overall environment tools rather than a specific knowledge base so I would actually consider rag to be a tool that an agent can have right so an agent

is able to have certain things like tools um but which I believe that there should be a slide for this um yeah actually these are I think I don't know if I have a slide specifically for what I mean when it comes to Tool use um okay yeah so I'll I'll talk about this like I'll talk about this a little bit later but um so the way that I would see it is not necessarily that rag is not an agent in itself you can think about an agent as the ability of doing something right like

agents are usually said as something like a marketing agent or a coding or software engineer agent it kind of embodies a specific task or a specific role rag which is like being able to reference certain information from a database is more of a tool that an agent can use in order to accomplish that task does that make sense actually tool use is the second agentic design pattern so agents can leverage external tools like web search or code execution to enhance its capabilities so for example you might want to get a research agent to uh you

wanted to buy a coffee machine right but instead of just asking it to figure out like what's the best coffee machine possible you can give it web search capabilities and compile the search results so you're letting it have um access to the internet and then it would be able to compile the search results and give you a better answer as opposed to if you're just directly questioning it you can also see like if it's a researcher uh agent and you want it to be able to have a rag abilities so it's able to reference like

certain data databases uh that it has access to you can ask it something like according to this specific database like what is it that I can do so that would also be part of like a tool that an agent would have the ability to use another example will be making a coding application so if you want to make a website do some math or like do some object detection whatever it is that you wanted to do you can give the agent D tool in order to execute code then and to be able to write code

so then it's able to do these computations and actually come up with a final result does that make sense see comments okay what is rag so sorry I should actually specify so rag is retrieval augmented generation it's a way um that you're able to give agents um it's it's a way for you to give the agent or whatever it is that you're doing uh access to a certain database does that make sense so for example if you're asking um if you're just asking AI something generic like oh like can you um I don't know come

up with God I can't think right now come up with a graph that is emphas that a graph that showcases revenue forecast for this specific grocery train right so it might be able to search the internet or it might just be able to have that information and it's able to come up with a graph but say if you want to ask a question like hey can you come up with a graph to show the forecasting for a certain grocery chain this specific grocery chain and you give it if you provide a specific data that that

it wouldn't be able to find elsewhere then that agent where that AI is going to be able to use rag to retrieve the information from the database that you're provided to give you better results I hope that made sense any other questions kind of okay um do you think they already have ai like Jarvis or very very Advanced like fully autonomous AI privately no no I don't think so I I think that commercial use without getting into politics like I I would very much doubt if that were the case is reflection part of reasoning um

I mean reasoning it is is like a reflection is a part of reason you're basically prompting the AI to do reflection which is a form of reasoning there's a lot of ways to get AI to reason you can give it like different prompts you can do different things but think about it in the context of Agents specifically um you're asking it to reflect upon itself remember like agents is defined by that circular process right like circular process uh and you're simply like trying to figure out how do you get it to reflect as part of

that circular process um uh basically give some contextual information exactly so that's what rag is you're providing it the ability to have access to data that it otherwise would not have to be able to reference that data to give you better results that's pretty much it like proprietary information for example can it do web scraping return data as Json yes it can so that's part of the tool use that is here right like if you have an agent like a gentic design pattern and you're giving it ability to do tool use use for example execute

code and do web scraping so that's part of um if you're giving the tool to do that then it would be able to do that and return data as Json this is something that has already been shown to work for like many times now so this is something that definitely does work um that'll be another tool that your agent can use exactly correct okay hopefully that made sense so I think the confusion that you may have it's like oh like why are we talking about these like specific agentic design patterns like what actually is an

agent and I think that can confusion is very valid just because there we are not we have not actually achieved an autonomous agent yet so we're really just thinking about ways how do we get better results it's just an evolution of how to get better results by using a framework in terms of getting a AI to embody like a certain role to do a certain thing and giving it tools and giving it different ways to be able to do that better I hope that makes sense okay so the next agentic design pattern is planning and

reasoning which is you know what we talked about just now um I mentioned this previously so this a gentic design pattern is asking the AI to determine optimal steps and select tools that execute a task efficiently for example you may be able to ask a um AI agent to generate an image where a girl is reading a book and her pose is the same as the boy in the example then describe the new image with your voice so first you the agent is going to be is going to need to be like okay like that's

like a lot right so he needs to be able to plan and reason like what are the actual stepbystep things that it needs to do in order to come up with a final um in order to describe it with the voice right so it has to go through like a sequence of different uh plans and different components what's the right word it needs to go through like a sequence of different steps in order to get to the final result at the end um so in this case you would be generating an image where an AI

where a girl is reading a book so first it has to be like okay so I have like this picture which is maybe like example. jpeg I need to be able to identify the pose I then need to figure out how to generate images and then I need to make the image into text and then I have to go get a tool that uses text to speech conversion so it's kind of like a coming up with a step-by-step process and this is a very very powerful agentic design pattern because you're basically giving your agent the

ability to plan ahead of things and you're giving it the ability to execute on that plan as well which is very similar to humans you know and in some ways like agents what we're the way that we're thinking about it right now is like how do we get it closer to the way that humans operate and this is something that gets us one step further okay before I go into multi-agent systems which is actually like really really cool I'm going to see if there's any questions that anybody has yeah like Chain of Thought prompting like

all these prompting methodologies is are me like planning methodologies these like prompt engineering things are ways in order to get a agents to do certain things well like design patterns agentic design patterns are different Frameworks that you can use in order to make an agent better BR force keeps scaling the AI up and pray I mean if if the whole like deep seek situation tells teaches us anything is the fact that maybe now it's not the time just to brute force and just pile on money and resources anymore there are like smarter ways of doing

things and I think AI agents as a framework is one of those solutions for how is it that we can think about using AI to accomplish our task in a way that's actually a lot smarter than just brute forcing things kind of like a shovel can't take a hole without you behind it AI can be a team of shovels for yourself yep exactly um use a reward function for every successful staff and action it takes well that's something that's already like part of the infrastructure of how AI is being created all right let us move

on so the final design pattern is going to be multi-agent design pattern so a multi multiple AI models with specialized roles so this is interesting because initially like all the other design patterns we talked about we're just talking about like one specific agent right that you're giving it tools you're you're getting it to think about stuff you're getting it to plan and reason and break things down to steps but this one's really interesting like multi-agent systems instead of just having one agent you are actually asking um you're you're actually asking something but then you're going

to get different agents that have different specialized task in order to accomplish something so why is it that this is this is useful or interesting right it it's think about human dynamics right if you go ask a human just like some person and go like hey like make me a marketing campaign then that single person has to be a jack of all traits like it has to be able to um you know research the market he needs to come up with like a concept he needs to do all the designs the drawings all that that's

like a lot to expect from a single human to do and it's actually the same thing as an agent if you just ask like one agent to do everything for you it's you wouldn't actually get that good results from it it's better just like a human team to have different um humans that specialize in specific things you can have market research humans so humans that only do market research and then that team you have people who only do design you have people who only come up uh who do implementation you have people who only do

report generation you have people who only do like the data analysis for things and all of that together is going to be a much better result um than if you just ask a single person to go do a marketing campaign does that make sense that's why like people started getting into multi-agent systems realizing mimicking human dynamics realizing that hey like if you can get AI agents to have specific tasks and specific roles and tell them to just basically like put them in a room and and then just go like work together you would actually get

better results than just trying to get one AI agent to do everything for you another example here is say you're writing a research paper you can have a research agent that would gather relevant papers and data you have analysis agent that interprets the findings and identifies patterns you have a writing agent that would draft the paper and you also have an editor agent that would check for accuracy and Clarity and finally you have a citation agent that ensures proper referencing this is going to be much better result than just being like oh you know write

this whole thing for me and make sure that it's done properly um yeah and then the actual Dynamics between like different multi-agent systems is something else that also becomes like a really like a really really interesting problem because you're essentially telling you know you're essentially being like Oh like there's a bunch of Agents together work with each other to figure out the specific task so just like humans there's a lot of research that is currently being done like what's the right Dynamic are you supposed to have like one manager agent that's supposed to be in

charge like a bunch of like you know like um yeah like one manage your agent and have a hierarchy are is it supposed to be like more like a sequential kind of situation are you giving the agents freedom in order to do whatever they want or you are should you like specify specific agents do certain things so once you have like a bunch of these agents running around and doing stuff you can start getting like pretty chaotic but but when it actually works well you can also get some really cool results as well and I

actually go into a little bit in terms of the current state of affairs like kind of like what people are working on right now and the different patterns that are emerging depending on the specific problem they're trying to solve so I will go into that in just a little bit um but let us see if anybody has any questions so far um nice to see you stream I thought I was banned I haven't seen a stream N I was just taking a very long break that was that was pretty much it but now I feel

rejuvenated I totally burned myself out when I was doing like a live stream a week plus making a video a week plus you know run lonely octopus the company that was a lot so I was like can't and I didn't want to like produce subpar content right so I just took a break but I am back now they talk to each other yeah yeah yeah what that Meme it's just like talk to each other work with each other here A bunch of AI agents y'all figure it out together um crew Ai and storm will be

example exactly exactly so multi um agent Frameworks like crei um autogen like these are all like examples of Frameworks of multi-agent design patterns which is really really cool and I actually show you uh I did the course in C basically did all of the courses on all of them and tried all of them and I want to show you guys like the ones from crew aai the way that they categorize it I would say crew AI is probably the best in terms of creating a multi-agent system that can actually get you to do something assuming

that you know how to code if you don't um yeah like there's ways of doing it as well but if you do know how to code crew AI I think is the way to go um is there going to be a replay yes do not worry there's going to be a replay you can also have better output if agents are able to make and complete sub goals exactly yes so that's why that's the thing right you get an agent to have a more specific task just like how you tell people to do something more specific

they're usually going to be do a better job than if you tell them to do something very very general it's like do this like you know do this entire thing write this entire campaign right this entire research paper it's you're not going to get as good of a result as opposed if you just tell to do like very specific things instead Tina the sign up still isn't working what is not working for the workshop signups like you can't sign up to it let me double check from my end does anybody else have problems with signing

up for the link that I just showed you we were having a lot of problems with our website previously it works for me yeah can can can can everybody else check as well like does this work for you guys please let me know if it doesn't yeah okay cool all right oh also yeah we ask you uh I think you just reply with octopus when you get the email so that it doesn't go to spam okay all right let's talk about multi-agent Design multi-agent Systems so I also oh yeah so remember if you can't remember

uh what are the four different agentic design patterns like this helps a lot just think about this pneumonic red turtles paint murals reflection tool use planning and multi-agents all right remember that red turtles paint murals okay so the next one is I also have a pneumonic for this for multi-agent Design Systems um specifically this one is going to be for like what are the components that make up an agent which is tiop Paka mixed tea I don't know does this help anyone by the way like for me it's like it's just easier for me to

remember like thinking about prompt engineering Frameworks as well like I have a pneumonics for all of them because I cannot remember like the specifics for things um anyway so like now let's actually talk about like what are the components of a singular agent before we start you know putting them together and telling them to to do stuff oh I was not sharing my screen that's awkward okay all right so multi-agent so tired alpacas mixed heed okay these are the components that are of a single agent the different components that you have and you need to

have these things before you can start mixing them all together okay so first you need to have a task so for example if you have like a travel planning agent like a singular agent you you need to have a task right for example it could be a plan a 3-day Tokyo budget trip for two people and then you need to have an output like an answer to from your agent so it can give you answers like daily itinerary hotel bookings actively activity reservation cost breakdown and transportation details and you have to have a model so

you need to like actually have a model model that which model do you want to use uh it can be claw it can be like you know Gemini whatever uh llama deep seek whatever you want you can trade that out as well and then finally you need to give your AI agent um tools so in example if you're doing a travel planning agent some of the tools would be Google Maps API booking.com API flight search weather data and Curr and currency converter so what I think is like really really cool is this like framework you

can trade these things out so these are like the basic basic components of all AI agents but you can trade out what the task is what the answer you want from it what the model is and what the tool is and I just think like at least for me like when I when when I was learning this like this helps so much because then I understand like okay these are the components the building blocks of what an agent is and that's that's literally it so yeah can't remember T alaka mixt is the pneumonic task answer

model and tools all right okay so now let's talk about I somebody does anybody have any questions about the components um of an agent before we talk about putting them all together um well yeah you haven't received the slides yet because I'm currently still going through the slides after the live stream we'll send you guys the slides why Al Paco and mixed tea I don't know it's just a pneumonic isn't it like isn't that easier to remember okay if I told you if I told you just to remember all right guys task answer model and

tools can you remember that but if I told you to remember this tired of pacus mix T then you can remember this phrase pretty easily and then you can think oh what is T what is tired Paca mix t t is going to be task a is going to be answers mix is going to m is going to be model and T is going to be tools right does that make sense that's why pneumonics are useful I mean you don't have to remember the pneumonic if you can just like somehow remember this directly that works

too task answer models and tools yeah you can you can you can do that too no problem um T answer is My Trivia what what language should I focus on just python I python is going to be the easiest one to focus on um there's also like some I'll show some like I'll share some no code versions of this as well it's the no code versions are actually pretty damn good so even if you don't want to do something with code um I think you can get away with that uh I would say it's if

you really want to create something that is going to be useful and it can be something that's incorporate it into a business and have like actual business use case and not be really really expensive that I still think you should use something like crew aai in which it's going to be like you you're still going to need to know how to code but if you just want to build something for yourself where where you don't mind paying more money to do something and maybe it's like not perfect like you can't get it to be super

custom there are um different no code versions of creating agents as well um do you need different prompts aligned against different tools or API for agent yes great question that is a great question yes you do you do that's actually something I did not include in this frame maybe I should change this framework so you do need a prompt that is absolutely correct so this is the components that you give it but the prompting part is going to be the task right the when you're writing what the task is that you want the agent to

do um this is where the prompting comes in so all the prompt engineering stuff I think I did a video a while back maybe I'll do like a refresher course for you guys at some point through live stream let me know if you want me to do that but that's where the prompt engineering part comes in the way in which you are prompting an AI agent every AI agent when you're giving a task you need to give it a prompt as well so you pre-load these prompts and modify them based on the agent Behavior outcomes

not exactly so what you do there are pre-made versions but when you're creating an agent it's better to just actually design it yourself so say think about it like if you're designing a video game character what you want to do is like the different components of it is going to be these four components you want to give your video game your character a task to do so if it's a travel planning agent you want to give it the task of planning travel right and then um you wanted to tell it what kind of answer it

is that you want so travel output you wanted to have daily itinerary hotel bookings whatever and then you want to assign a model to it like which model you want your character to have maybe you wanted to have a claw model and then you also assign what tools you're giving the agent so that it's allowed to use in order to accomplish this task so you want to set this upfront when you're designing your character when you're designing your agent so that it's able to do that specific task and then you can iterate on this like

maybe for example you realize that your agent like needs another tool so that you can add in another tool for your agent as well maybe you realize that maybe like cloud is like not not the best model for some reason and you want to use an open source model you can switch out a model as well make sense any other questions absolutely refresh your course on prompting or prompt and sharing would be great okay we can do that I can do that how long approximate is today's live stream it's going to be approximately an hour

um I don't know if I'm going to have time to go through like the live Dems and stuff but I might do that as part two if you guys want yeah I hope it's I hope you guys don't mind that like do you like this kind of more theoretical kind of uh that I actually give you the fundamentals of things as opposed to just like directly showing you how to do things like do you like that I give you Frameworks and stuff that's like a question that I actually would be interested in knowing if you

actually like that or not um and actually speaking of which if you guys can do me a favor this would be genuinely really really helpful um if you could so what we're trying to do more marketing research over here and also come up with better live like what is it called different like workshops so [Music] uh if you guys fill out this form honestly this would be super helpful so it's asking for like testimonials in terms of like what it is that um like feedback and testimonials like that would genuinely be super helpful let me

just kind of like share this page because what we're trying to design like what we actually want the workshops to look like it's very helpful to be able to come up with with like what people actually want so it's kind of like a for you to provide feedback like this is what it looks like um yeah it's really simple it's just your name um and then how did the live stream help you what did you find the most valuable and if you want you can have a video testimonial which would help me a lot so

you know having those videos and showcasing people hey our workshops are actually helpful um yeah that' be really great if y'all can do that for me all right let us continue you okay okay let me see what people are saying yes I would like refresh your impr prompt engineering what agents have you built I've built quite a few a lot of them are just more like personal ones and we started building things that relate to like content process and things like that cuz that's the uh I've been trying to incorporate more business aspects into it

framework first and then example would be helpful okay yeah okay okay cool yeah cuz that's what I was thinking I'm just like oh like is it better if I just show people how to do stuff because most that's what most YouTubers do right they kind of just go like oh like this is how you build the ultimate system for creating things and I'm like I could do that but for me personally like when I have Frameworks and pneumonics like actual Frameworks for what are the components of an AI agent it helps a lot more than

just like somebody just showing me things um okay so it seems like you guys do like the fundamental try it's like not super annoying that I don't just show you exactly how to do things um okay cool ther I just could not gladly willing to provide feedback what is the current examples of agentic AI and Ai and AI agents uh yeah so some of the examples like for example an AI agent here that was saying earlier travel planning agent this is this is an example of what an agent would have right it would have the

task of planning a 3-day Tokyo budget trip for two people it would have an answer in terms of what you wanted to do it has a model and it has a tool so this would be an example of an agent whatever you think about like a human role that somebody does just add the word agent and then you'll have an agent version of that super simp like travel planner agent you have a travel planner you have a travel planner agent you have a software engineer you have a software engineer agent yeah this is like literally

the easiest ways to figure out what an agent can do like theoretically can do just think about what is the human version of that and then there's going to be an agent version of that we don't know how good it's going to be I'm not saying that it would be as good as a human version it's going to vary depending on what it is but the whole idea of an agent is that it's supposed to be something that can accomplish a task or like have the specific rule that a human would all right I'm going

to talk about multi-agent designs design patterns before we run out of time okay so the first one is sequential pattern so you can get different agents and you ask it to basically be in an assembly line so each agent will complete a task before passing on to the next agent for example you can have a document um uh if you wanted to create a task that's scanning documents extracting information and coming up with an action plan at the end right so you have agent one and agent one's job is going to scan documents so what

would that actually look like going back to this framework right um task answer model and tools like agent one since it's its job is to scan a document like from a PDF or something the task would be scan documents answer that you wanted to have is a scanned version of the document so you would have U maybe one as a Json maybe one as a text file you know whatever it is that you want to ultimately have a version of the text from the document the model you can give it you can choose Claud if

you like um my general default by the way is usually clawed for for most of these things these days and then the tool that you're going to need to give it would be like Vision right because it's scanning a document you would need to give it the ability of accessing camera or accessing some type of vision in order to scan the document and they would also probably need to have like give it ability if you're wanted to create a Json or something like that have the ability for like code where like certain I guess like

you don't really need it to have code you can just reformat it um as that but it definitely would need to have the ability of scanning information um either through like the camera app or like some other type of API and then you would pass it on to agent two which is supposed to extract the text from that information so it's basically just like a bunch of things you're extracting it and then you need to extract the text from the information that's there oh sorry I think in this case we're actually breaking it down sorry

about that let me try again in this case we're actually breaking that into two different agents the first one all is going to do is scan the document so it would only scan the document and then it would only have the tool of um having camera for example right and then it would pass that on to agent two who then would extract the text from the scan document so agent two would actually have the task of extracting the text and you would give it a certain model and specific tools as well and agent three so

after extracted text it would pass it on to agent three which would then summarize the content from the extracted text into something that makes sense and then you would pass it on to agent four which would create an action items based upon the documents that's being scanned does that make sense so it's kind of like an assembly line one agent does something passes on to the next agent it passes on to the next agent it passes on to the agent after that this is a design pattern a multi-agent design pattern that is really useful um

for things that require a strict order of operations like if you can clearly Define what are the steps in order to accomplish a task then it's this is a really good design pattern for that let me see if there's any comments conceptual understanding helps a lot okay cool I'm really glad I'm really glad this structure is great okay yeah if you can provide like feedback in terms of the thing I just sent which I will just send to you guys again that genuinely is like extremely extremely helpful um to figure out because there's like 335

of you guys watching right now and I don't know what you guys are thinking so got let me know um all right so the next one is going to be H hierar patter I just do not know how to pronounce this word so in Hier Aral design pattern is when you have like one agent um at the at the top which is going to be a manager agent and then you have a bunch of Minion agents or sorry employee agents who would be doing specific task and then they would be reporting it back to the

manager so the manager would be assigning a task and it would be coordinating work um and delegating to these different sub agents and then the sub agents sorry sub agents is better than minions yes let's called them sub agents um the sub agents will do specific things report back to the manager and then the manager would be able to compile everything together so an example of this is if you're trying to create a business report the manager agent would oversee the report creation and it would assign its sub agents sub agent number one to do

market research sub agent number two to do like financial analysis and maybe sub agent number three it would tell us to do competitor analysis right and each of these agents would again come with Ty toaca mixed heed it's going to come with its own task it's going to come with its own answers its own model and it's also going to have its own tools um so yeah that then after it does all these things it would report back to the manager agent and then ultimately the manager agent would have an output which is going to

be the business report so this is quite helpful that is for complex projects that need coordination so um hierarchal is useful when it's not necessarily like an assembly line it's just that there's different components that need to be done and put together it's generally better if you just assign it to different sub agents make sense any questions can we make an AI agent hedge fund I don't see why not I think that would be a hierarchical kind of situation right minions plus one minions sounds too much like slavery uh should I just say minions then

you delegate to your NPC minions all the things that you don't want to do okay so nice I have parallel system multi-agent design pattern by the way isn't this like kind of like at least for me when I first learned about these multi-agent design patterns was like literally blowing my mind isn't it so cool that you're basically getting a bunch of Agents like humans and just tell them to like figure things out themselves and work together as a unit I I don't know I I thought it was it was really like blowing my mind when

I was understanding this um okay all right let's talk about um the next one over here which is a parallel system so parallel system is multiple agents that are working simultaneously and there's no dependency between task and an example of this would be if you want to do like some sort of data analysis um and you the agent one is going to be analyzing sales data agent two is going to be processing customer feedback and agent three is going to be reviewing market trends so they don't necessarily have to come together um from a business

perspective like each of them is just doing kind of doing its own thing and all of them are necessary in order to ultimately be able to run the business uh and this is really helpful for large scale data processing so this is actually quite similar to the way that um from the code perspective like a technical perspective how we do a lot of data processing to begin with and instead of like sequentially executing like one line of code and things like that and processing data you can actually split the data into like different parts right

and then tell a different person or a different software or you know different Hardware to compute at specific things does that make sense so you're like say a really easy example would be like say you have I don't know like 10,000 lines of of data right instead of having to process all of that you can actually just split that up and then tell each one to just process them by themselves individually and this is also kind of like task as well so each agent will be doing like a specific task when you're doing data analysis

um in parallel make sense any questions comments thoughts very interesting it would be nice to follow along to build some time-saving projects would be useful in my case an agent that helps as a virtual assistant for Knowledge Management for applying for jobs yeah I think I don't think I have time to do that now but we can do like a practical implementation of this I actually show you how to do this next time if you guys want where is AI lacking in contrast to where it wants to be oh my God it's lacking so much

okay that's a that's like a whole other topic there's a lot to be desired right now I think we're honestly like we're in such infancy at this point um there's a lot okay the next design pattern is asynchronous is an asynchronous design pattern so you can have like agents working in dependently without fixed timings and this would be more just reactive to events and conditions so a great example of this is like in cyber security you have different agents that are just doing different things at different times and all working completely independently which is why

it's called asynchronous so agent one could be monitoring Network traffic agent two could be checking system logs and agent three could be analyzing user Behavior right and then for each of these agents they're all checking for cyber security threats in their specific domain and there's specific um this one is checking for Network traffic any anomalies this one is checking for system log anomalies and this one is checking for user Behavior anomalies and if any of these agents as they're doing things independently detects an anomaly it would kind of flag that right so this is quite

useful for real-time monitoring and response any questions thoughts what happens when when an agent glitches into the team effort that's that's actually a great question so there's ways of getting around that so that's part of like the key the the magic of prompting so you can actually specify when you're specifying agent in the prompt like what it is that you need it to do in case something happens right so you can kind of like have constraints you can kind of catch certain things that happen um a really great way of doing this is the answer

component for what you want the agent to answer you want it to be like a very concrete answer so as opposed to having free flowing textt you can tell it to output something like a Json and then you can actually check the Json that's coming in to make sure that it's correct before passing on to another agent so there's a lot that you need to play around with in terms of the way that each agent is able to Output certain things and what what each agent is doing and that is just magic of prompt engineering

mostly and trial and error a lot of times um so do the sub agent have their own UI or do they work in background at their own independent API that honestly depends um generally speaking agents they don't necessarily need to have a UI like as long as it's it accomplishes the task that it's supposed to accomplish um so each agent has followed tired AP Paka mixt correct exactly yay you guys like my pneumonics I'm slowly getting into your head next time you think about what are the components of making an AI agent tired alaka mix

tea correct every single agent needs to have a task it needs to have an answer it needs to have a model and it needs to have a tool do you guys remember the first one so I'm like if there's two I guess my goal for this entire thing right now this um this entire Workshop is two acronyms all right let's do like a little quiz now two acronyms red red turtles paint murals and tiop Paca mixed tea Pop Quiz what do those two things mean if you remember those two pneumonics that means you have absorbed

the information I have taught you so far all right pop quiz time do you [Music] remember what's another use case for this besides cyber security good question so for asynchronous design patterns let me think let me know if you guys can think of anything something that doesn't require that's like just independently doing something completely by itself H I'm thinking like any sort of monitoring system like if you have camera like um what is that called like CCTV networks right like each camera um like a lot of security stuff I can see that cuz they're just

trying to detect any anomalies right I can also see it being used for or yeah like anything that involves like security and monitoring would be would do something like that what else h o let me ask AI while you guys are answering the two questions what do they stand for okay all right type in the chat I'mma quiz you guys tired what does tired alpacas make tea stand for what does red turtles paint murals stand for [Music] okay task answer model yay you guys are learning things you can remember okay looks great for detect Ah

that's a yeah that's a great one like detecting oh that's a really cool one like in healthcare system I can totally see that like I know that again not like a healthcare expert here but it's like I think you're doing like different scans and trying to like figure out what's wrong with someone right you're like conducting a bunch of different tests and if you have like a different agent that's responsible for different test test and you're just getting the results and interpreting that information that can be like an asynchronous design pattern as well okay so

I don't know if I have time to go through the next component because I do got to go I I do have to do something afterwards but I want to give you like a little taste for what it is that we're going to do next um so I did create this um practial so the N n8n if you want to check it out is a platform it's a that you can do a no code solution to creating agents so in this case like this is the this is what it looks like it's like a flow

where you're using Telegram in order to communicate with the AI and this specific AI agent it's using open ai's model so again Ty pacas Ty pacas make te so the task that it has is that I wanted to I wanted it to be able to take information like my daily what it is I have to do and output a schedule for me and then eventually also put that into my Google Calendar that's the task that I gave it um the answer is I wanted it to first of all using telegram to confirm um what it

is that I wanted to do and then also be able to come up with the calendar events and then schedule the calendar events into my Google Calendar as well um make te so m is going to be the model so I used open AI model in this case and then T is going uh in terms of tools so it had tools the tools that I had is I had the ability to access my Google Calendar um and it also had the ability of creating calendar events yeah damn I don't know if I have time to

get through this right now let me see if I can like 1 minute okay maybe this is going to have to be a next time situation I'm going to have to end on a cliffhanger to actually show you guys what it looks like we'd be mad at me if if I ended it on a cliffhanger and actually like show you guys what it looks like also by the way I am making a video so a video tonight is going to drop where I'm going to be covering this this is kind of like a snippet there

in the video there's going to be even more information that covers this as well so you should check it out um yeah and I do show the demo in that video but to actual implementation I might have to end out a cliffhanger and show you guys next time we can build something next time okay see if anybody has any comments no Cliffhanger Cliffhanger okay hey okay great people are answering the question where is going to show up next time all right show up next time all right and then we'll actually show you guys how it

is how we can build this um can you suggest they no Co MTI agents platform okay yeah yeah how about this I'll give I'll give you guys some homework for people who are going to do no code stuff what I'm going to tell you to do is go play around with this platform uh hold on it is honestly quite good and they have a free trial so you don't need to pay I forgot how long like a week or two weeks or something like that so you can get started for free using this so homework

time so oh my God I have too many tabs open too many tabs open okay so no code platform and then if you do know how to code I highly recommend using crew AI so I'm going to send you the link for for that as well I mean you can use both if you like but depending on what it is that you want to do code version you can use crei cool what would the next section be sign up for I will let you know sign up for the workshop and we'll give you details for

the next session so this one and by the way like again this one is if you this link is going to be if you um want to get the slides so I will like literally after this live stream I will send you guys the slides for this so you can go through it um and then if you can give me feedback and testimonials for how you find a workshop that would actually be really helpful as well so I'm just going to give you the link to that as well so I know what it is that

people actually want to do also yes cool all right if there were 321 people standing in front of you would you like would you do this this so comfortably I think I'm good at pretending to be comt comfortable all right thank you so much for joining and I will see you guys next time I hope this was helpful and don't forget to sign up so you get the slides okay see you bye