Episode 5 : Agentic Core

Key Takeaways

And so far we have seen three three four episodes. So we are building the application agentic professional network. So so far we launched uh few features profiles and connections. So in this episode we did the foundation yesterday. I'm going to show how overall agentic core is going to work. So let me start sharing my screen. Yeah. So episode 4 is focused on agentic core. I'm going to talk about a agents uh teach with an application so that you get the context. How do you build agentic core for a particular application? And in this case I'm taking agentic network which is nothing but a professional networking application I'm reimagining and I'll generate synthetic data uh so that I can test uh to discover the profile based on the goals that is set and I'll do some testing. Okay. So before going into the details I want to take a slide. How do I reim reimagine the agentic core for professional networking application? As soon as a person signs up set up a profile. Uh the profile agent should activate and start generating the summary suggest the headline identify the missing profile section and then make it complete. And as soon as that is done, so another agent should start um sending request or receive the request, analyze it and gatekeeper agent should screen the request and recommend whether it should accept we should uh the person should accept reject or review based on the goals of the person whether he's whether he is willing to build a networking with the skills or whether is uh is interested to expand And that industry connects based on the self evaluation by AI. What is the score relevance and what is the confidence so that a person can just accept or reject based on a suggestion and trust agent can run in the background to build your reputation. So building a brand personal brand is very tough many people does not do. So trust agent will help uh to improve. So this is the thought process. So I spent only 1 hour yesterday uh after I closed episode uh 4 uh and then I'm continuing in this episode I basically I also spend one or two hours only that I get uh but the speed at which the AI such as cloud CLI etc. So it's very rapid now for development. Okay. So now let me stop sharing the slides uh and then uh start sharing the actual code. Yeah. So what I did uh I made some changes uh after yesterday. Uh as I mentioned I given the specification uh the builder network uh specification uh where is it so I given the specification And then I let it run uh because this is an epic so it need to do a lot of work. Uh so I let it run around 30 minutes. Uh and then I did some conversational uh coding. So uh for 30 minutes more to tweak little bit etc. So now after all the code is generated so I deployed. So let us check what are the things it deployed and I cleaned the database. I think we'll start the testing again fresh. So it deployed uh so many functions. Okay. So these are the functions it deployed and the hosting is updated. So let me uh start registering to user. Okay. So now let me sign in. Yeah, I have fixed that glitch in the login also. So now uh this is the first profile. Okay. So after creation of this it should create the summary etc. Okay. So why it took agent is configured? Let's see. By default it's making agentic. Yeah. So it enable certain agents. So that it is agent first approach. If you want to change, you can change. So step three. So now I want to set industry connection. So let's set uh different um goals for this profile. Let me set industry networking and connections. Target industry let's say technology target tools let's say engineer can select more than one uh okay so these are the uh basically uh input for a agents to uh personalize your experience. So if you have any uh particular um company names, you can put it here. So let me uh okay, we'll come back to that later. Okay. Now let me launch agent. Okay. Now let me just search check uh okay that agent activity will be shown here or not. Give me a minute. I'll just set it up. Uh yeah. So now I expect say this is the one I'm developing incrementally. So now the profile agent should kick in as per our plan. So any activity should come here. Let me ask lad uh whether it is going to show agent activities. ideally you should see agent actions one more collection. So whether it started or not we'll check uh because it coded so there may be some uh bug uh so we'll just check uh okay so ideally the profile agent should uh kick in Um, that is our intention. In the meantime, let's fix this first and we can create uh we can have two sessions also. Let this figure out whether agent uh activities are shown in the front end. And parally let me also ask uh okay uh I created a profile user show agent profile profile agent kicked in Correct. So now once you achieve the maturity of working with one cloud CLA session then you can slowly improve kind of having more than one session. Okay. It's like uh writing one horse and uh once you kind of get a expert in writing one horse then you can kind of try to uh learn how to write multiple horses together. So right now I'm trying to do one uh clot CLI session to figure out uh okay why uh the agentic uh is the agentic activ agent activities are going to shown in this space in the UI another agent um uh clot serialization so now it figured out based on the analysis the answer is no the dashboard currently does not display the agent activities so it's trying to code so it has already put a to-do actions and it's trying to do and other uh session I'm trying to figure out whether the profile agent kicked in. Uh so now it is saying I found the trigger there is an on create trigger and uh profile agent store the data in agent actions. Let's see whether that profile uh agent actions is stored. No, we don't see agent actions. Ideally, it should go check if it ran. Go to file. Okay. Yeah, I don't see option to go to the log of on profile created. Okay. So now let's see that also. So these are the functions which is nothing but a cloud function on profile create if you want to see the log. So you can see the logs here which will take you to the Google cloud log console. So there are no logs. Okay. Let's uh tell these two information to uh yeah see say now there is nothing um there is no standard prompts that you learn and then do. So if somebody is saying that I think they are cheating you. So you need to kind of you need to dynamically think what you need to ask. So in the earlier stages people are saying the next generation language is English. So now you need to frame that prompt yourself uh based on the situation. So in this case uh profile agent did not trigger. So first I asked okay can you please uh help me to find out is it working? Then it asked me to check these two. So I checked and then I framed the next prompt. So nobody is going to tell you okay these are the set of prompt which you can take and work okay you need to start thinking how do we do prompt engineering and dynamically start thinking uh these are the things uh you need to learn how to do the ways of working with AI. Uh so then AI will guide you to do the next step. So now it is asking permission uh to check whether that function is properly exported uh etc. So it will start reaging. Okay, let it triage and then we'll go check the other one. So now what is saying? Okay, now I have updated the dashboard to display. Okay, so now the front end logic to retrieve from the back end agent action is updated. So now we'll deploy this function. Okay, now it is only front end. Uh so I'll go to front end npm run build and deploy front end change can happen. But uh one more page is going on. I found the issue profile creation happens correctly. Since there are no logs at all, mostly causes functions are not deployed. Okay. So it is asking me to do check etc. So I'm if I give permission to run that command, it will run that command in a bash uh process and then it will continue the triage uh because it has a capability to do all of this. So we don't need to just follow the instructions what A is telling. We can also give the instructions to AI to do the more work. Sometimes it will be more cautious. Uh sometimes it may need permission. Sometime it will be uh kind of asking you to do. So now I'm asking um yeah so now it is asking first it's setting up the Firebase use. So it is starting to run the commands. So Firebase use make sure the right project is uh selected in this case agentic network. So now let me give permission for uh running uh fireplace commands. So don't ask me every time. So you need to kind of balance it out when to give permission, when you don't want to give permission. It's all whether you want to run in gear one and gear two or gear three. So if you want to run in gear 4 uh or gear 5, if you want to run at a speed of 130 km/h, you cannot uh kind of uh keep too much control. You need to give full control to that and then give clear guidance uh specification then it can run faster. So I'm giving full if to check the log every time don't ask permission I've given that. So now it will start checking the log. So I don't need to go fetch the log and give it. It will find the 30 lines of uh log and then it will analyze by itself. Yeah, this is interesting. So once it knows how to do uh right once you give permission then your development typically developer doesn't check the logs or if you need to check the log you need to go to the server uh if needed after development and deployment if things doesn't work it will go check the logs and then uh based on the uh intent identified in the error it will try to make the changes. So that is what it's trying to do because the function did not work. Yeah. So now this is the root cause. So so interesting to see uh there are different models will behave differently. So it found out finally we are using GPT5 uh mini and GPT5 mini does not support a temperature parameter. So and now it is trying to switch it to 4.2. Sometimes it takes a decision uh I don't want to switch to 4.2 remain in 5 mini and then remove temperature is what I need to tell AI to make changes but it quickly trashed and then found out the root cause. keep using GPT 5 mini. Uh whenever whenever you make calls to GPT5 mini please avoid sending temperature s but however when you call when you make other models use temperature. So as you do programming you will know about characteristic of different models. So this one I figured it out I think sometime back. So GPT5 mini does not take temperature. It's a small model and it works uh based on certain characteristic. So now it is uh changing a code based on the guidance. If it is uh not GPT 5 mini uh then uh GPT models without temperature. Okay. Again it is taking GPT5. If when you want to give other instruction you need to do a escape and then tell. So right now I'm going to give instruction use only for GPD5 mini not for GP5. Okay. So this is like uh say now when you go or drive a car say sometimes you follow the map uh say if you need to take a left or right you take a decision. So if AI says you need to take a right and uh but uh logically map is wrong and you may need to take a left right. So you need to guide that uh sometimes otherwise you need to if you take completely follow AI you may end up an issue end up on an issue but you you will finally eventually correct that but you'll take little longer cycle so now you need to watch closely what A is doing and then you need to get the work done Uh yeah. So now it updated the code. Say look at the beauty of it. It is just deploying that particular function which is got impacted which is updated. So now since we have given the permission to do a Firebase commands. So it is uh deploying only the function on profile created manually. Sometimes I am little lazy. I compile everything and deploy all the function. It is unnecessary effort. Okay, let me delete this uh profile. So that when we create next time it should start creating so basically whatever I'm teaching uh say the A you need to build application. So that is very important because every time you will face a different situation. So we'll keep two UI one is this uh one more incognito we put Okay, now this user is deleted. Ideally, that session should not be there. Yeah, it came out. Okay. Okay. So, it deployed H. So I think it is running update function it triggered. Okay. So now let us wait for a minute and then uh start creating a profile. Okay. So let me create a profile. Yeah, content profile gets too quickly. Yeah. Let's see the logs. Yeah. So now new profile created for the user. Activating profile agent. So now the agent started uh working. So let's see what it does. Finished with the status. Okay. Uh completeness score 50. Summary confidence 65 suggestion zero. So let's see any agent activities has come database. So agent actions. So so now one document has come. Initial profile analysis has done. So the completeness reasoning and summary uh okay now it is updated or not. Okay. Okay, now let's go to view profile back to dashboard edit profile. Okay, so it made progress. So now let's see uh I think now let me ask that this summary should have been updated in uh see it created a nice summary. Uh okay so which I did not write. Uh so now let us see ask that uh again uh okay now wait I'll go ID it should have been updated in the profile and then it should have reflected when you see view profile and edit profile so let's ask that in view profile it is not there edit profile it's not there okay let's ask that Yeah. Yeah. Great. Uh now I see and also here it should come. Uh, I think I lost that. Yeah. So now this is the next step. So it's incrementally development uh right. So before you take the requirement and you go and code manually. So now you need to do prompt first approach uh and conversational approach. So now it created uh the trigger and it uh fired the profile agent. The profile agent got the agent work completed. Uh so now that agent work need to go back and get updated into the profile and then whatever agent did that agent activity should come. So that I don't want to manually code. Uh so uh the root cause approval page please. Okay. So now agent actions are going into user ID agent action and approval Q reads from okay update store the profile analysis in approval Q and suggestions Make necessary changes. So you can manually code and build agentic application. Number two, you can leverage AI to build application. Now what I am doing is combining both of that. Uh I'm leveraging AI to build application and that application is agentic application. So yeah, that's what we are doing. Yeah, now it went into plan mode because now we are asking a design question, right? Uh yeah, so see the intelligence now based on the root cause analysis fixing an issue or based on the prompt if it is a design it is changing the act. So now there are two functionality one the agent activity and approval. So now if you see uh network so a agents you have pending approvals. So now oh okay it has come here h approval. It is given analysis pending. So I did not see it here. Okay. So now this approval is done. All right. I think we need to add it back button. Okay. Yeah. So there are functions I think maybe it is trying to figure it out. So now that summary has come here. So now I don't want only summary should be there. Uh okay we'll update it here. Ideally should come like this. It would have come like this. So we'll tweak that. Uh right now it's trying to do ending review summary. Okay, let's Come on. Yeah, things are starting towards say building an application typically you do follow different process before say so you do product backlog sprint backlog uh analying requirement and then doing the design and then you do development so it's all dynamic now I think it's changing to different agent activity And you need to be careful once you ask the question it will take the action. It will never say I will not do etc. So you need to be very careful uh and you need to get the maturity u when to ask and uh how to ask etc. when we create the second profile. I think these changes will do. Let's not delete this profile. We see and in the meantime I'll just show you uh right last time yesterday uh we saw the dashboard. So this is the dashboard. Now it gave uh the connection functions. also my connections there are no connections what are the requests came what are the uh requests sent so these are the features which were built yesterday and then um you can discover people uh there are no profile once you create the profile it will allow you to discover and then uh it gives the suggestion so once you have the profile so then based on the goals that you set uh it will give a a based suggestions for your network and any incoming request or send request. So those functions were added yesterday and a agents uh you can configure and change the goals and pending up rules. These are the three functions are added. So this is our connections basically networks feature uh module and this is the AI uh management modules and I just added beta because I'm deleting the database uh users are suddenly if somebody use this uh application and registered as a user so I might clean up that uh because I'm doing the build of this product uh gradually sometimes I may need to clean. I'm not going to migrate those users. So, uh whoever is registering so they are kind of trial uh uh users only. So see agent actions change to agent activities. Now it's updating the activity card of the dashboard And uh as you work with command prompt command CLA so you will be well versed with few options. If you see control T control T it will show the to-do list. So interf will show you the thinking process. So now you will you will kind of once you familiarize you will know all of these uh shortcuts and they're also introducing more features. So uh I think it's a is a is a way uh to develop in the future because you don't need an ID uh ID you look at one file at a time and uh you write one character at a time uh all that is changing and even if you do ID so you take uh if you want to do gear one gear two gear one approach if you want to make contextual changes within one file. I do it in IDE opening that particular file and asking AI. Uh there also it will pull the context across the project. It's not only one file but CLA mode you don't need to bother about that. It will find out the context uh quickly by running some grip commands and uh understanding the code base uh which file to read uh more than one files and then what are the files to be changed impacted analysis etc. So slowly people will move from ID based development to CLA based development which was the way Unix developers did before uh they used to use a command u etc which are more powerful uh so now I think it's coming back uh with a yeah now it's deploying the changes uh both for the two functions and also I think it should deploy UI also let's Okay. see if you see it worked continuously for 5 minutes. So within 5 minutes it planned within 5 minutes it made changes it compiled and it is deploying. So that 5 minutes can be longer also based on the what is the work uh that is being done. It's Yeah. So it is completed. So now yeah, I think we need to create a user again. Uh so now let me just ask. Okay. What is the old one? Aging actions. Okay, let me clean up. I think with this I uh we'll create a new profile. Maybe I need to create a automated uh user creation. Next time I'll do it. appro. Why is it outside? I think this also should move. I think I need to put some rule somewhere. It is every time is creating some collection outside the user. So this I think it kept outside. I'm moving it up. And here I'm asking uh just to save only the summary. So uh ideally when you develop uh this right I recommend uh say many of the developers now I have uh now they have big screen so better to have 13 screen and then you divide your layout into four the one I shown in episode 1 and two uh so that will make you to see all the context at a time and then uh you don't need to switch uh so you can watch multiple context and and uh give commands. I'm not doing multiscreen because uh in YouTube video many people are seeing in laptops. Uh so that's why I'm restricting myself to show in laptop. But ideally my developer experience increase uh if I do uh split the screen into four and keep uh one or two command line prompt and uh one or two browser windows um so that I can see all of them. I don't need to switch to see what is happening. But when you do parallel work, I think both are conflicting then it will become problem. Uh so when you do parallel work so choose the isolated boundary. Uh right now I think this is changing the structure this is changing the function. So I think there may be a possibility of conflict. So when you do parallel actions on the same project uh so you need to be careful. But anyway, I started. Let us see. Yeah, don't do parallelism until you are expert. Okay. And otherwise it it will you will get into problem. So one of the task is completed. So here uh we are moving the direction. Let's see. Yeah. So it's almost done. And I think it's deploying the rules for permissions because we change the structure. So it knows if there is any structure changes I need to give fire base permission. So it moved to a user user ID and approval Q. So it is giving permission only for the user who are logged in that particular user. so it's deploying all the functions now. Maybe nowadays um there is another trend. So now the AI is capable of doing faster and your laptop also need to be upgraded. Uh so there are laptops with GPUs. So if you want to if you if the value is there I think uh you can upgrade your laptop to GPU laptops it can work faster. Yeah. So this validation is done. Uh we cleared the database. Yep. Yeah. See agent reputation. Agent reputation is the agent which works every hour to improve your reputation. I think this also should go under user. This is the last agent which we'll worry about. I give some command to work I think logically So with this one agent um with the two three created synthetic data I leave this work this isolated task. Yeah. See now we divided the task into that one prompt into four subtask. First it generates a synthetic data user profile CSV and create the fire store load script. Create fire store cleanup script and add npm commands. So logically that one prompt is divided into four subtask. Uh so typically you will create one user story and some four task and assign different task to different developers. So all of that is not needed. So based on the prompt it will identify what are the different task and now this is the task. Now let us see what it is thinking. So control O. Yeah. So control O it is generate doing the first task. First task is to generate synthetic data. So now it is thinking what user wants me to do. Uh and then it is planning out uh okay now what are the different industry, what are the different role skill etc etc. So this is the thinking. So if you want to switch between thinking and then task you need to switch control troll t for viewing the task uh other it will show the current task current task uh it will show which is generating a synthetic user profile but there are four task in the pipeline and uh for that particular task if you do control it will show what that clot CLA is thinking about it and then uh it is thinking and then it is uh got some variant. After that it is thinking and then it is decided to put the script under function script and then it is thinking then it is creating a typescript uh to generate synthetic. So if you want to know what's happening uh behind the scene you can do control uh t control. Okay now it is deployed. So let's create two profiles. Sign in. Yeah. Let me not write too much. menu is going to be updated by AI. Okay. Okay. Now, okay. It is asking. I don't need I think so. We'll go to the clear. Your user is created. User activity is created. Now that activity should come here. Okay, maybe it is autonomous. It will see. Yeah. So now it is an autonomous um level is high. So it automatically did that update. It did not ask for an approval. But that activity should have been shown. Uh okay. Before it asked for an approval because it did not check the autonomous level. So here uh we had reasoning reasoning was removed. Uh that is good. But this bracket also need to be removed. Basically I want to get only this. No no it did not remove reasoning. Okay. So this will fix it. So but it did autonomous update. So let us check um agentic Okay. Now what happened to synthetic data? We'll see. So after that we'll do uh some discovery and suggestion then we'll close this session. Yeah. So now builder network functions scripts. Okay. Generating. Deeply conscious. Okay. Then I know this one too. So when it all when I allowed it to run it figured out there is a error and it is fixing that error. So if I run manually I get an error. I need to copy paste this and then type it here. So if you allow it to run it will figure out what is the error and it will fix that. Uh so let it fix and parallely I'm doing asking the fix for uh displaying the agent activity here. So once our synthetic data has come we will uh look at more um network suggestions etc. Right now there are no more data. So it is not ending any suggestion. Yeah. So now it has put some fix for uh synthe data generation. So I'm running it. Now it is also checking okay is agent log is written uh whether the previous dependency is working okay now it has updated from indexes so we need to see still agent it has not come let's see any error in the browser once you know where it is also. Agent Yeah. We're getting the synthetic data. Okay. Now it generated the file. Yeah. It's loading the synthetic gate. See it generated synthetic data for 34 industries, 40 locations across the globe uh 100 plus skills uh 30 networking goes say the diversity of the test data you see what and all is generated. Now it is asking me to run it. So before running it will just fix this first. In the meantime, I'll open the synthetic data. See? Yeah. So all is in CSV format. So around thousands of users are uh generated. So we'll load that and check Don't know what to say. Agent activity created created is there. Okay. Come on. So I'm getting some error like check for the synthetic data load generation did not load. Okay. So now The conjur And then deploy I think we'll close at uh in a minute. So maybe we'll continue this u later because it's already one and a half hour. We'll just check logically if agentic um agents activity is showing and uh is uh synthetic data is uploaded and search is working. project settings or The present motion of the pain. H. So now it uh got to agent activities. So two pending approvals. So it has shown the agent activities the profile update summary. So when you click here you will uh see these two agent uh approval. So once you approve uh then it will do the actions actually it will show the confidence also. So based on this you can either approve or reject. So this action should work. Looks like some terrorist. Yeah, it loaded thousand profiles. So, let us see that. Okay, now it loaded 1 th00and profiles. Okay, let's see that. Uh, yeah. So now it generated synthetic users. So you cannot login with those users but you can use that for discovery. So now let's go to the one minute. We'll just see. Okay, let it come. network discover profiles. Yeah. So now the thousand profiles are there. Either you can manually discover uh right. Let's look at um yeah cyber security. Right. So it's coming. Uh maybe we'll test that functionality separately. Uh so let me do some search. Yeah. So I think it is working. Maybe we'll look at deeper on the search functionality. So I did a search on Salesforce. So it is given three uh profile. So you can do a connect So now we have sent one request. So you will see that in send request. So now let's see uh AI is suggesting uh so now when you do AI suggestions so these are personalized recommendation by uh a uh agents. So since my profile is a empowered engineer, it is suggesting MLOps engineer and it is also suggesting few more uh connections so that we can tweak in uh how the agent should work. So now the test data has come. Uh now we need to build the feature uh by testing it. Uh that we'll do it in um next episode. Right now uh the agent uh the profile agent is firing in and agent activities are coming here and uh a generated summary has come and synthetic data we generated we uploaded. So now I think we'll logically conclude here uh we'll just fix the top part and then we'll conclude. Yeah. So now for the approval I think it has put some fix. So it figured out there is a m mismatch with the user might have documents created with the new code. Okay. So I think it is putting some fix. I think we'll ignore that for now. So what we targeted uh it took little longer uh because we are doing live version. It is not I created something and I am teaching. So I'm showing you this because this is the way you need to learn uh let's say otherwise you will not learn say when you do things how you need to do prompt and how you need to build an application rapidly this that is what I'm teaching in this episodes. So now we talked about changing gears we built user stories and feature and epic in yesterday's session and today's session. So now we'll build on more uh right now the basic um profile creation improved with the networking. So now you can do a connections uh whatever sent request uh uh everything you should see that will test it and gradually we'll strengthen the connections networking uh module the send request should come here the one we sent and a agents are there. So the basic capabilities for networking connections is there and basic capabilities of agents are there. So in the coming episodes uh we will strengthen each of the agent and each of the functions of this application. Yeah. So this is uh fifth episode within five episodes and spending half an hour and 1 hour we could build an application of this level. uh so that is the power u that you can leverage with the tools such as cloud CLI to build your application faster when you build application don't build in traditional way make sure you have agentic core and also don't go overboard on multi- aent um the buzzwords of agentic uh world you need to build the functional capability and leverage agentic as a core where it is needed. You need to manage between deterministic system. Balance between deterministic and nondeterministic and leverage AI where it is powerful and then ensure context engineering and data engineering to make a agents powerful. Uh so those things we'll talk about in detail in next episode. Yeah. Thanks everyone for watching. Have a good day.