Full transcript of “Generative AI For Testers | What is AI Agent? How It Works? | LLM Vs AI Agents”. Click a timestamp to open the video at that moment. Need another video? Use the YouTube transcript generator.
0:02 hey everyone welcome you all so in the today's video we will try to understand about AI agents so we know about LLMs large language models like charge GBT or deepseek or Google Gemini these are all different type of LLM applications where we can provide different type of prompts and we can get different type of outputs but what exactly AI agent means how exactly it will work and what is the difference between LLMs and AI agents and where exactly the AI agents will be used so we'll try to understand in today's video so before knowing about AI agents we need to understand what are the limitations in LLM all kinds of LLMs so the LLMs are basically trained on data up to a
0:51 certain point and cannot access real time or updated information so for example let's say charge GPT is trained uh data up to 2021 but now if I give any prompt if I ask charg to provide any information which is happened after 2021 or it may happened around 20 2022 or 20 2022 or 3 or 4 then charge GBT may not provide you the accurate information so why because charge GPT is trained data up to 2021 so that is the biggest limitation similarly if I ask LLM or if you ask any charge GP or any LLM application about a recently released software tool or recently
1:37 happened event it may not provide you the accurate outcome so that is one of the biggest limitation in any LLM so basically it works based on NLP so natural language processing so it can understand our prompt and accordingly whatever the data which is already pre predefined inside it and accordingly it will provide the output that's how the LLM works so the lack of realtime knowledge it doesn't have any knowledge about real time that is one of the biggest limitation in LMS similarly no contextual memory if I take any LLM so it doesn't maintain any memory of past interactions suppose today I have interacted with LLMs and provided different type of prompts right each prompt will
2:27 be treated as a different or independent query or independent question so it cannot maintain all these uh prompts related tracking it doesn't have any tracking and it doesn't maintain any memory inside it so it doesn't have any contextual memory so inability to perform actions and LLMs basically provide the outcome based upon the prompt but it cannot perform any actions so again what is action means for example if I ask any LLM what is the output of 2 + 2 it may gives you right result which is four right that is also an action but what action I'm talking about here is the realtime
3:13 action suppose if I ask LLM to perform some action on some e-commerce website so for example I want to connect to some e-commerce application then I want to search for some product and I want add to the cart and then I want to do some payment so this is my actions so if I ask to perform those actions in LLM it cannot perform so because it is nowhere related to realtime actions or realtime activities so whatever the content which is already having which is more static and based on that only it will provide the output for us so it is inability to perform the actions llms cannot perform the live actions suppose if I want to interact with an application I want to perform
4:00 certain activities on the web UI so our chart GPT or any TC cannot perform directly actions on the web UI right it cannot interact with the backend systems basically it cannot connect to any database or any API or any other resources it works based on the static content so that is the reason it cannot perform the realtime actions and the other limitation is bias and insecur and inaccuracy sometimes even LLM may provide some uh inaccurate information or inaccurate outcome whenever you ask some prompt sometimes it can provide the inaccurate outcome that means it can bias the process so LLM may produce a biased or incorrect outputs based on their
4:49 training data so suppose if I ask something and the related data is not available or not uh feed into the LLM so you may get bypass to bias to some other content and it can provide that content so that is a possibility which we have in LLMs and finally limited understanding so LLMs can't truly understand the context like human suppose if I talk to some other person they can understand what exactly you are asking for and your intentions and your behavior and everything but LLMs cannot do that so they just generate the responses based on the patterns not actual comprehension so these are all limitations which we can identify in any type of LLMs so to
5:37 overcome these limitations we go for AI agent so what exactly AI agent means so a AI agent is a program or it is a system so that uses artificial intelligence to perform a task autonomously so which is very very important term autonomously means what without our instructions or without our our human intervention AI agent itself can do the actions in real time that means what AI agents can directly interact with the backend systems it can be database or it can be web UI or it can be API whatever resources are available at the back end our AA agent will able to communicate with them
6:27 and perform the realtime actions so as a user I provide the prompt what a AI agent will do it can understand that prompt and it will perform the realtime actions but in LLM the realtime actions are not possible whereas in AI agent realtime actions are possible so according to the definition so an AI agent is a system that uses artificial intelligence to perform task autonomously it can perceive its environment so to perform task autonomously means in the live or in the real-time actions to perform it persuas its environment it will able to take the definition uh decisions and also take actions to achieve specific goals so we need to define what exactly
7:17 the agent AI agent should do so according to our instructions the a AI agent will perform the task at the real time that means it is able to interact with the web pages it is able to identify the elements automatically and it is able to perform the actions right according to our interactions it will able to do everything so to do those things it will internally maintain the environment and based on our prompt and based upon our previous uh interactions it will identify the history based on that it will make the decisions and accordingly it will take the actions based upon spec our goals so AJ agents often combine LLMs with other tools APIs and decision-m algorithms but how exactly a agents
8:06 works it is combined with LLMs it is combined with LM so LLM is always provide only static data but sometimes AI agents also can be integrated with LLMs so then LLM is capable of doing the realtime actions also AI agents can integrate with some other tools or sometimes it can integrate with APIs and it can also integrate with some decision- making algorithms so that's how the A agent works so AI agents is a kind of a enhancement of LLMs we can say so LLM works based on only static data which is predefined or pre-trained whereas AA agents will work on top of LLMs
8:55 where the AI agents will interact with the live environment or real applications or realtime data so that's how the AI agent works the main features autonomy so autonomy is nothing but what we can operate without constant human intervention so we don't need suppose if you just want to work with LLM so we need to keep giving different prompts depends on the outcome but once you provided the instructions to the AA agent it will able to perform all the actions constantly without human intervention that is called autonomy adaptability means what it learns and improves over time suppose if keep on interacting with the AI agent it learns what exactly you're asking right it
9:44 learns from that and improves over the time and once you started using AI agents you can keep on getting the best results and goal oriented it is designed to achieve the specific objectives so why a AI agent design because we need to achieve the specific objective so for example let's say a customer support AI agent is there and what exactly a customer support AI agent will do it can answer the queries resolve the issues and escalate problems to human agents if it is needed so most of the e-commerce applications we will able to see that so whenever you start charting in the chat board so artificial intelligence AI agent will invoke so the AI agent will able to answer your queries and suppose if you are working on e-commerce application if you want to know the
10:35 details about the product you can interact with the AI agent which is related to that application even it seems like a chatboard at the back end AI agent is integrated so whatever queries you are asking about the product it will go internally it will communicate and it will capture the data in deeper and that will consolidate the data and provide the answers to your queries and if there are any issues about your orders right so your AA agent will able to resolve them and sometimes if the A agent is not able to resource it is automatically escalate the same issues to the human agents if it is needed so that's how exactly the agent works another example for example as a
11:20 chat board so that not only responds to user queries but also books a ticket for a flight so for example Some banking applications some traveling applications you may notice this so there will be a chatboard will be open once you successfully log in and in the chatboard it will ask your details user profile and all the details then you can just book the ticket uh flight and everything directly from the chatboard so instead of interacting with all the different things or to do different workflows instead of doing multiple navigations and workflows you can directly achieve that through chat box itself and that is internally integrated with AI agent so simply it will you can just provide the query like I want to book a ticket then it will
12:06 ask you so where is the source and destination and date so once you provided that information automatically the ticket will be booked or available options will be shown so then once we selected the option the ticket will be booked so all this process will be taken care by AI agent okay so that's how exactly the AA agent works if I look at here this picture we can understand uh in deeper aspect so this is basically a agent let's say the user as a user we provide simple a prompt and the a agent in use NLP natural language processing to understand the prompt and generate the task list and then it will able to execute and while executing the task it needs some resources
12:53 it can be uh data document or it can be code executor or it can be some task output or any uh machine learning model or any LLM so this Agent will be integrated with any of these sources so whenever the user is gives the prompt to uh the A agent will do the task or will complete the task or execute the task by using a different tools and environments so how a agents overcome the limitation LM so we have understood what is LLM what is AI agent but how come the AI agent will overcome the limitations of LLM like AI agents can integrate with the tools so for example AI agents can connect to APIs databases or tools to perform the actions but LLMs cannot connect them so LLMs
13:46 directly cannot connect to the backend systems it can be database or APIs anything so it provides only the it works only based on the static data so whenever you provide the point it provide the static content or static outcome but once AI agent is integrated with this LLM then that AI agent will able to connect corresponding APIs or databases or tools then it will able to perform the actions example as we discussed booking a ticket or fetching a realtime data can be done through AI agent and memory and context so if I use only LLM it doesn't maintain any memory or it doesn't know about any context it is just provide the outcome based upon the pre-trained data but
14:31 once you integrate an AI agent the AI agents can maintain the context across interactions by storing conversation history or using external memory systems so whenever you provide different prompts and based on that it can understand what exactly the context is and it can store that conversation history in external memory systems it will internally maintain the memory also and realtime data access as we discussed AI agents will directly interact with the real-time data accordingly it will perform the realtime actions but LLMs cannot directly interact with the real time but once integrated with AI agents that will directly interact with the uh live or realtime data and accordingly it will perform the actions the realtime data in the sense
15:22 fetch the live data from the internet or any other sources to provide up toate informations LLM cannot provide the data up to date right so it can always provide the old content similarly uh it's not a a agent is not like that a agent will always pull the data from the real systems or real databases or live applications so real time data access is possible through a agent task automation is possible task automation suppose if you want to do a single task or multiple tasks we can automate that using a agent so just we simply provide the instructions ai agent will perform those tasks the automate the task so they can perform automated task unlike LLMs so LLMs cannot
16:10 perform any automation just whatever the prompt we provided accordingly it will provide the outcome but it cannot perform any automation but once you provide proper instructions or prompts to the A agent will able to perform the automation which only provide the text responses llms will provide only text responses they can execute task like sending mail so the uh a agents is able to execute the task like sending emails automatically updating the spreadsheets controlling smart devices these are all realtime actions so normal LLMs cannot do this it can just provide the static text responses but once integrated AI agents will able to do all these task so by integrating this
16:57 LLM generative AI and AI agents there are so many automation tools are coming into market which are basically codeless automation tools we don't need to write any automation script or code just we need to provide English uh statements just simply provide the test case in normal plain English and accordingly it will perform all the actions on the web UI or applications example let's say test trigger so test trigger test trigger is a tool uh or we can other words we can say it is a a agent internally used uh we don't need to write any coding inside this we just provide the plain English language uh steps and accordingly it will perform all the actions on the real applications okay later we will see how exactly they will work so this is all about AI agents what is LLM what
17:49 are the limitations of LLM how exactly a agents will overcome come the limitations of LLM so as I said an example an A agent can use an LLM to understand the user request like suppose the A agent is integrated with LLM and we are providing some prompt then what what is the prompt like say book a flight to New York and then connect to the travel API so as soon as you provide this prompt to LLM internally the AI agent is integrated with this LLM so what this AI agent will do it will connect to the travel API corresponding travel API and then perform the booking process okay so where exactly we are providing the prompt on LLM but internally the A agent is able to communicate
18:36 with the realtime data that can be database that can be API or any other source so that it will perform the realtime activity so what is the main difference between LMA agent so far we understood so LLM is basically having a pre-trained data it is not up to date static data is maintained it doesn't have any our context it cannot perform any execution or any actions on the real applications or real time whereas a agent will do the task it can remember the context accordingly it will perform the actions and execution on the realtime data so if I look at LLM what is the main role of LLM is it can simply generate the test cases if I talk with respect to test testing software testing
19:24 basically LLMs can be generated test cases llm can generate the code also right or document test results based on the natural language inputs and what is the limitation here is they cannot execute the test it can generate the test cases it can do all the documentation stuff but it cannot execute the test on the real applications or it cannot interact with the software system directly it cannot identify elements it cannot perform the actions on buttons input labels or any logos or anything right so it can suggest test scenarios we can write test cases we can prepare the test plans all the documentation work we can do that but we cannot execute it we cannot run it for that we need to go for a agent so A agents can do that job so A agents can automate the
20:14 entire testing process from generating test cases to executing them and reporting bugs everything we able to do by a agent what are the capabilities they can interact with the software as I said it directly interact with the application simulate the user actions like what exactly we do on the applications let's say login we enter username password submit then we will verify the dashboard page is opened or not so all these interactions or actions can be done by AI agents and also it can integrate with CI/CD pipelines all these processes we can achieve only through AI agent example an AI agent can automatically test a web application by clicking buttons filling forms and validating
21:01 outputs so here exactly the AI agent come into picture so AI agents will be integrated with multiple things basically LLMs will integrated with AI agent so overall what we can understand about AI agent is let's say this is our LLM so large language model so we will provide certain prompt here as a user we provide the prompt so LLM will have only static content so it is having only pre-trained data so from this it will give the response it cannot interact with the live applications suppose if you are integrated agent AI agent so then what we can provide here we can
21:48 also ask LM to provide the actions or execution so it can directly interact with the A agent and this AA agent will able to perform the actions realtime data it can access it can be database it can be API it can be web applications so all these can things can be accessed by AI agent and that will be integrated with LLMs so nowadays there are so many automation tools like codeless automation tools like test rigger will internally use AI agent through which just we need to write the test cases in plain English language and that will be executed on the realtime browser it will able to interact with the API it can access the database it can perform the web testing it can perform API testing it can perform the mobile app testing all things are possible
22:34 with the AI agent okay so that's how we need to understand about AI agents in the coming videos we will try to understand more about AI agents that's all for this video thanks for watching
