{
  "video_id": "QVawxdMRZtk",
  "channel_slug": "deeplearningai",
  "channel_handle": "DeepLearningAI",
  "title": "AI Dev 26 x SF | Eda Zhou & Mahdi Ghodsi: Building Personal AI Agents with Open Source Models",
  "duration_seconds": 2025,
  "url": "https://www.youtube.com/watch?v=QVawxdMRZtk",
  "upload_date": "20260520",
  "transcript": "So today we're here for a little bit of\na different kind of talk. So this one is\ngoing to be more of a a hands-on\nworkshop. So the first few slides going\nto be by my coworker here and uh then\nI'll take over the workshop part. That's\na hands-on part. So my name is Maddie.\nSo we're both from AMD and we prepared\num basically more of how you guys can go\nfrom an open source model, deploy your\nown model on a GPU, connect that to your\nopen claw and you know build something\ntogether. Uh so\nyeah, if you want to participate in the\nhands-on aspect of the workshop, make\nsure you're signed up for AMD developer\nprogram. We have GPUs available for you\nto follow along live during the\nworkshop. And we also have custom\nlobster squishies uh for folks who\nfinish a challenge.\nHey, those are AMD branded lobsters, so\nyou don't want to miss them. Um,\nand before we start, just a show of\nhands for me to know how many of you\nguys have ever used AMD GPUs for your AI\nworkload.\nI see a few hands. Uh, a lot of\npotential though. So, today will be your\nchance. So we have limited number of\nGPUs that we're going to share the link\nwith you guys later. Uh you can log in\nand access that notebook. So it is a\nchance for you guys to take out your\nlaptop if you have it with you. We do\nhave about 120 instances of dedicated\nGPUs for you guys that you can access.\nSo while go through the slides, maybe\nit's good time for you to take off your\nlaptop, connect to Wi-Fi, and be ready\nfor the second part.\nI pass it to\nThank you, Motti.\nAll right. Uh, so if you want to follow\neither of us on LinkedIn, here are the\nQR codes. Um, this is my alter ego as a\nlobster.\nAnd so, real quick, uh, what is an LLM?\nSo, we're just going to quickly go\nthrough the foundations here. So, an LLM\nis a model trained on a large amount of\ndata. It predicts the next token and by\nitself, it's just a text generation\nsystem. So, it can answer questions. It\ncan um generate and explain your code,\nbut what it can't do is it can't act on\nexternal systems. It doesn't maintain\nits own persistent memory or manage a\nmulti-step process. So on its own, an\nLLM just generates a response. It can't\nindependently do these other tasks. And\nso this is the gap that we're going to\nfocus on today and try and fill.\nCan I get the speaker notes?\nLouder.\nThe speaker notes.\nOkay. And so as we step through it,\nwe're going to step through filling this\ngap. So like we said, an LLM is trained\nto answer questions to predict the next\ntoken. A chatbot then has the\nconversation history injected into it.\nSo it feels like it's remembering your\nconversation, but really that's just\nadded into the prompt and it sees all of\nthat in the input. Now the agent is\nwhere the next the real big step\nhappens. So agent is where you start\nadding memory planning ability to call\nAPIs uh take actions into loop until a\ntask is done.\nAnd so really important part of agents\nis the loop. So the react loop stands\nfor reason and action. And it's actually\nthree parts. So reason, action, and\nobservation.\nSo with a loop um the agent can\ncontinuously think through a problem\nuntil it solves it. So if you think you\nask a chatbot, hey, can you fix this\nproblem for me? It'll give you a\nresponse, but you don't actually know\nwhether that response is true or not,\nwhether it works or not. With an agent,\nit'll actually call a tool. Maybe it'll\ncall the code, call some test code, and\nit'll test if its fix actually works. If\nit doesn't, it'll go back and try\nsomething else. and it'll keep going\nthrough that loop until it verifies that\nthe problem is solved.\nSo with this loop, the base model uh is\nimportant, but actually a stronger loop\ncould make your agent process even\nstronger.\nSo there's three components to building\nan agent. Uh there's the model, the\nruntime, and the tools. And so breaking\nit down, model folks are probably\nfamiliar with here. This is the large\nlanguage model. Uh there's examples of\nreally popular recent models and all of\nthese had day zero support on AMD\nInstinct GPUs.\nSo the runtime is very important. It\nmanages the loop context retries the\nstate. Uh these are the guard rails.\nThey uh know when to make the tool calls\nand then catch failures as well and to\ncontinue executing that loop.\nFinally are the tools themselves. These\nare those external capabilities. So, MCP\nservers, uh, APIs, CLIs, and these are\nactually usually abstracted from the\nmodel itself so that you can use a lot\nof different kinds of tools. It doesn't\nneed to know those details. And MCP\nservers, there's plenty of open servers.\nUh, if you've been to Mott's previous\nworkshops, you'll know, you're familiar\nwith the Airbnb one. Uh, there's GitHub\nones. Uh, there's plenty of ones that\nyou can just use right out of the box.\nAnd so finally, how is Open Claw\ndifferent than just these agent\nframeworks? So a lot of agent frameworks\nare just libraries. You import them,\nwrite loops. Uh they're reactive, so\nthey wait for you, and they only run\nwhen you run the script. Open Claw is\npersistent and is an application. And so\nit's persistently running on your\nmachine. And so if you want a cron job\nor a heartbeat monitor, it's just an\ninput to that application. Whereas if\nit's a framework approach, you need to\nbuild a scheduler around it. You need to\nbuild more uh stuff around it in order\nfor a cron job to actually work. And so\nthe open claw approach, it just works\nbecause that's how the architecture is\nstructured.\nAnd so without further ado, uh if you\nwant to participate in the hands-on\nportion, uh you need to go to this URL.\nTake out your laptops, go to the URL,\nand it will start spinning up your\nworkshop. We'll give everyone some time\nto get this set up, and then Marty will\ngo through it live.\nYeah, we'll leave this up for a couple\nminutes. So, please, if you have your\nlaptop, log in. As I mentioned, we have\nabout 125 dedicated MI uh 325 available\nfor you guys that you can actually go\ndeploy the model, then connect that\nmodel to your open claw agent that we\nbuilt together today. Uh it is kind of a\nfirst first come first serve. So if you\nhave your laptop, I believe not all of\nyou guys have laptops. So a good portion\nof you guys will be able to access this\ntoday. Uh so we leave it up a couple\nminutes and then I'll switch to the same\nnotebook that you guys going to be\naccessing today. So if you guys um have\nnever done any uh model deployment on\nGPU, that's actually another question\nfor me just to know how many of you guys\nhave ever deployed a model uh on your\nown GPU or hardware before.\nOkay, that's a lot more. That's great.\nUm, so to do so, you usually need some\nsort of framework or a serving framework\nto do so. So today, we're going to use\nVLM.\nSo VLM is a very uh popular framework\nright now made for um GPUs without you\nknow restriction to any specific brand.\nSo uh we're using VLM as one of the most\npopular serving frameworks. So to do so\nyou will have access to you know this\nnotebook that has all the instructions\navailable for you guys. If I want to\nstart this uh model so right now we're\nusing 3.520B\nI can just copy this which is going here\non the terminal and paste it here.\nSo what is this doing right now is as\nsimple as VLM serve model name. The\nmodel name comes directly from hugging\nface. So, if I just go here on hoging\nface and look up models and put Quinn 3\nuh.520B\nuh this one right here. So, you can see\nit's actually the exact same same name\nthat I got from here. I could just copy\nand paste this uh right as the model\nname. And that's all you need basically\njust BLM serve model name. And this\nworks on AMD GPU without needing to do\nanything. It's just BLM as a first class\ncitizen on AMD hardware. So you can use\nthat. The rest of them are optional\nparameters I'm showing here to enable\ntool calling which is required for\nagents to be able to actually call the\ntools. You do need to have the extra\nparameters such as a tool call parser\nwhich comes directly again from the\nmodel provider. So it's a model based\nand API key is for the software such as\nopen cloud to you know add that level of\nuh security. they usually need you to\nadd the API key and that's for that\npurpose. The port number again is in for\nus to have as an optional added all\nthose and that's it. So once you run\nthat it will start to go and load the\nmodel weights on your GPU. Here we have\na AMD GPU uh and it will basically uh\nload the model the model weights uh go\nthrough all the VLM serving uh\nrequirements that we have and once it's\nready you're going to get you know the\nthree lines of hey the model is ready\nfor you while this is happening I will\ngo ahead and run this uh next cell this\nnext cell is all it's doing is just\nchecking to see if the server is ready\nso it's nothing important I don't want\nto spend time on this but we will know\nif the if the server is going to be\nMeanwhile, why this is happening, we'll\nget to open claw parts. So, Open Claw um\nbasically if you guys never used it,\nit's a very personable agent. You\ninstall it on your laptop or your\nmachine and it will have access to all\nyour files. We're not going to talk\nabout security part of it. you know\nthat's that's for them to figure out but\nit does have a lot of power and access\nand um you know you can disable any sort\nof access you want but it starts with a\nlot of power uh such as your files bash\nbash commands you can give it access to\nyour emails or whatever you want but you\ncan always you know go back and say I\ndon't want you to access that so to to\nyou to start openclaw they made a really\nnice uh openclaw onboard u command for\nyou guys that you start with so I was\nhoping you guys can do it. I I see maybe\na few of you guys may be doing this, but\nuh feel free to stop by and we can show\nyou how to do it on your own uh time.\nSo, on OpenCloud onboard basically uh\nwalks you through all the parameters you\nneed to set and it's an interactive u\nsession. It will take a couple seconds\nfor you guys to load. Uh but meanwhile,\nwe can also check to see if the model is\nready. It's already loaded but and the\nmodel is ready. So this is the server\ncheck I had\nand this is basically the VLM server\ncommand we ran. So that basically means\nthe model is ready for you guys to use.\nGoing back to uh what I had here for the\nopen cloud onboard\nfirst is basically accepting the risks\nas I mentioned we're not going to talk\nabout that today.\nAnd I'm gonna reorganize this so you\nguys can see the side by side.\nOkay. So except yes. Uh that will walk\nyou through the next part. So we have\nprovided all of this for you guys. So\nwhen you're on your own, you actually\nyou don't need me to tell you all of\nthis. It's all available for you guys to\ngo through. But we'll do it together\nhere. Uh so the model provider the\nreally cool thing about open-source\necosystem such as openclaw vlm sg lang\nis they're working well together and now\nuh with openclaw you don't need to\nactually go do any sort of you know\nweird configuration because it has it\navailable for you guys. So if you go\ndown all the way to here you should see\nvlm and uh other frameworks available\nsuch as sg lang which is another very\npopular uh serving framework. So, VM is\nthe one we chose. And you can see uh the\nport we chose 8,000. So, we're going to\nstick to that. The API key for\nsimplicity put APC 123\nand the model name is 3.52\nB.\nSo once you configure the model, that's\nbasically all you need. If you look at\nthe next couple of questions I have is\nall a skip and that will get you started\nimmediately to be able to use openclaw\nwith all of its power. But one thing to\nnote is uh open claw basically is as\npowerful as you make it. So the tools\nthat you give it access to is what makes\nopen claw do things for you. Um without\nthat is limited. It still works. We're\ngoing to see some of the examples but uh\nthat's for you guys to on your own time.\nAgain, if you want to set this up, spend\nsome time, see what kind of tools you\nwant to give it access to. So,\nbasically, I skipped all the questions\nof what channel, you could do, Telegram,\nyou could do WhatsApp, you could do all\nyour all your messaging apps. In terms\nof uh provider, uh we're going to skip\nfor now for uh any sort of search\nproviders they have. the skills I\nmentioned, we want to skip that for now,\nbut you can actually enable a lot of\nskills such as like do my Gmail access\nor um there's many other skills that you\ncan look up online uh that might be\nuseful for you guys, but uh you know\nGmail access, calendar access, uh web\nbrowsing, a lot of those are available\nfor you guys to choose and hash and\nterminal will make sure that our open\nclaw is up and running.\nOkay, the first time you run open claw,\nthis is kind of where it gets you. Um,\nit shows how personable it is. So, the\nway that works is the creator of open\nclaw created all these MD files for you\nguys. Soul.md is the soul of your agent.\nBasically, you tell it how to behave and\nit's really important and actually\ndepending on the model. I've had very\ninteresting experiences with soul.md.\nThere was a time that I I wanted to make\nmy agent to skip bunch of verifications\nbut in my solar MD I put do not believe\nanything do your own verification. So no\nmatter what I did it actually did not\naccept anything I wanted to do. It will\ngo back to solar MD that overrides any\nsort of rules that I give it. So this is\none of the most important files in your\nopen cloth.\nAgents MD is basically all the rules of\nhow you should behave, what you should\ndo and that comes uh packaged with open\ncloud. The identity MD is basically uh\nwho um you should be you kind of emojis\nthat you want. We're going to go through\nit right now and you're going to see how\nthat's going to be created. So while I\nleave this up uh I'll start interacting\nwith it and then we'll show again these\nfiles.\nThe first thing you should uh basically\ndo is introduction of yourself and\nintroduction of the agent to you. So\nhere for example it says hey I just came\nonline who am I who are you? So I'll\ncall it your rock roll and if you don't\nknow rockim is AMD's equivalent to CUDA.\nSo you're rock and roll. Uh my name is\nMaddie.\nYour vibe is uh rockstar.\nUh I'm in Pacific time zone.\nLet's start.\nSo what is going to do right now? It\nwill start overriding all these files I\nmentioned. So the soul MD, uh the\nidentity of MD, the user MD which is\nlike basically uh who I am uh any sort\nof tools that I wanted to give it. I\nbasically talk to it. I don't need to do\nany sort of configuration. That's one of\nthe really most powerful things about\nOpenClaw is even if it has a problem,\nyou don't need to go figure out what the\nproblem is. You tell it. You say, \"Hey,\nyou have a problem. I want you to\nbasically fix this problem.\" And it\nactually go and configures itself. It\nknows its own source code.\nSo again, it's kind of gone back. It\nwants to be very personable with me.\nIt's like what matters to you the most?\nYou know, any hard boundaries? anything\nthat you want me to, you know, take care\nof it for you. So, I'll say, you know,\nuh just be helpful.\nHelpful and don't ask many questions.\nAll right.\nSo once we answer these question that\nbootstrap.mmd which is basically initial\nmd file for it will be deleted and it'll\nbe set up for you based on what you\nactually ask it to do. Um and this is\nwhere I wanted to print some of these\nfiles that you see. So it says okay I\ndeleted my bootstrap and you know I'm\nready to roll for you and it has a vibe\nof you know that rockstar that I asked\nfor it and you can see all of those\nhere.\nSo basically based on what I answered\nit's no magic. It'll just keep notes of\nyour stuff. The software is really\nsmartly designed to be able to read this\nand always be available to you guys. So\nthe identity of MD is basically what you\nknow it should be called. Solid MD is\nwhat I want it to be like the vibe for\nit. The boundaries that I wanted to get\nthe vibe again uh\nwhat else? Ages MD is something that I\ndidn't change. We'll change that in a\nbit to show you how that works. But\nbasically that means it's all set up for\nyou guys. So that's for the initial\nsetup. Now let's let's show some\nexamples what we could do with this guy.\nSo agent MD as I mentioned is the rules\nfor the agent. I can ask it to be more\nof a you know a debugger for me for my\ncode. So if I have a buggy code I want\nsome sort of policy just don't invent\nanything. You can type all of that in MD\nor I can ask it. But I want to show for\ncompletion. You can do both. You can go\nmess around with agents. MD or you just\nbasically ask it to do. This one is the\none that you actually go mess around\nwith agent MD. So if I run this, it'll\nactually overwrite this in my agentmd\nwith some information that I want.\nUh so now it's ready for me to actually\nbe used for debugging. So for fun, what\nI had planned for you guys is um a\nlittle application that I made with a\nobvious bug in it. So this is basically\nan application on my GitHub. It's not on\nthis system and I want to show how easy\nit is for you guys to basically drop a\nGitHub people. If you have no coding\nexperience, if you don't know how to set\nup stuff, you don't need to. You could\nbasically just like drop it here and\nsay, \"Hey, uh, go figure out this\nproject, install it for me, and run it\nor like make it ready for me.\" So that's\nall I'm doing. And it will basically go\nlook at the GitHub repo, pull it on your\ncomputer, installs it for you, and tells\nyou how to run it. U, okay, so we got\nthis installed for us and actually tells\nus how to run it. It says just go to\nthis this folder here. It's ready for\nyou guys to use and I will do that. So\njust to show you guys oops it's hard\nwithout a mouse but we got it.\nSo basically if I go here\nand open terminal\ninstall it the instructions told me that\nall you have to do is just do something\nlike this python\nmain again copy pasting\nand let's do a simple 110.\nSo this is a word per minute counter. Um\nI made this like basically vibe coded\nall of this to see if you I can make an\napplication and uh I'm just going to\nlike type random stuff here.\nThe problem is\ndoes anyone see\nthere's a bug right? Um and I\nintentionally introduced a bug to show\nyou guys how you can basically get this\nto debug for you and fix the code.\nThere's a really obvious bug here. I got\na zero accuracy. That was the\nintentional bug I had for it. But just\nto show you guys how EDG is is for\nopenclaw which is you know it's not sold\nas a coding agent as a general agent can\nactually do that depending on the model\nyou use and another thing I've heard is\nlike well open source models are not as\ngood I mean depends how you define\nthey're good enough they're good for a\nlot of use cases such as the one that\nI'm showing you right now. So if I just\nlike b b b b b b b b b b b b b b b b b b\nb b b b b b b b b b b b b b b b b b b b\nb b basically say, oh there seems to be\na a bug with this uh open type tester I\nthink it's called it. Um can you check?\nSo basically we'll be able to look at it\nand try to like debug the the problem.\nAnd this is all again on an open source.\nUh,\nokay. I should copy paste that. It's\nlooking as a different bug for\nuh Yeah. No.\nOkay, let's do it.\nSo basically it should be able to go\nfigure out the the test files for this\nand see if there's you know anything\nthat is getting wrong and read the code\nand try to fix it\nwhich you can also if you find anything\nuseful for you that you need to instruct\nyour agent to do over and over this is a\nvery very uh simple use case that I'm\nshowing which is kind of struggling\nright now for me run the test\nand see the problem\nin\nin my application that you installed.\nAll right. So while this is figuring out\nso um\nbasically what you could do with uh\nopenclaw hopefully this gets it but what\nyou could do with openclaw is you say\nyou have a task that is very redundant\nright like for example go on this\nwebsite and go read this material and\nextract this and then go you know email\nthis to this person and do that. So you\nhave a checklist of stuff you need to do\num you can actually ask it to create a\nskill for you and that skill becomes\navailable every time that you want to\nuse it. So it actually uses the skills\nfor it for it that would save that and\nskills with MD. So the agent every time\nyou ask a similar thing for it, it's\nlike well I know what to do. I'll do it\nfor you. And that's basically what we\nwanted to show you guys here that uh\nwith this um sort of problems like this\none that let's see if it found it.\nOkay, so basically he's found some sort\nof problem with it. It says it was\nshowing 0% so 80% accuracy. So I know\nthe problem uh that now apps work\ncorrectly. So I can fix it for you guys.\nSo basically it did the it did what we\nwanted. It fixed the problem and now we\ncan say okay now go create a skill for\nthis. I have some instructions for it\nthat the skill I want to make sure that\nyou you build is doing one two three and\nyou know the instructions are up to you\nand you're trying to show how you can\nget it to create a skill and then reuse\nit later.\nSo I was just prompted it and given the\ninstructions you want. I can create the\nskill and um we'll test it in a second\nto to show you guys how it's working. Um\nfor completion again all of these are\nwithin the environment of open claw. So\nonce you install your open claw such as\nuh the one we did everything resides in\nthis openclaw folder and you can always\ngo check it. If you don't like some sort\nof policy you made you can always go\nmanually delete it. So now I created\nskill. We can check to see if it's\navailable.\nHere we go. I ran this and this is what\nit's doing. Showing okay under skills\npyest debugger is the name of the skill\nI created. This is what I will do and\nyou can use it anytime in future.\nSo for using the skill uh openclaw has a\nlot of ways to do it. You can you don't\nhave to like remember exactly what you\nthought. You can literally naturally\ntalk to it. But if you want to enforce\nit, you want to make sure that it's\ndefinitely doing this. It has a lot of\nshortcuts too. So the slash that you put\nwill give you a lot of options that you\ncan see is it's showing me right now\nlike help or gateway status. But if you\ntype skill, it says okay, type the name.\nSo I can just type the name of the skill\nand give it a new uh problem to solve\nand it should remember what it you know\nit learned and apply that to the same\nthe new project that has never seen the\ncode for. So what I've done here says\nskill use the name that I just used pi\ntest debugger to create a skill. Go\ncreate go just like look at this GitHub\nrepo. I didn't even tell it what this\nis. I didn't even tell like go clone it.\nIt's going to automatically go through\nthe loop that it showed to reason. It's\nlike what is this guy doing? He's giving\nme a GitHub repo with no instructions\nbut I can basically go figure out what\nit is based on the skill I have. So\nwhile this is again applying the skills\nuh that we we talked about to figure out\nthis this new problem\num one other really cool thing with u\nopenclaw is how it could be a multi-\naent system and do a lot of stuff with\nyou well the agents can communicate with\neach other in an environment recently I\nwas working on a different project this\nis for you know our internal use that\nhey uh whenever we run any models on a\nGPU you can actually increase how fast\nthat model can run by going down to the\nvery low-level codes but it's a very\ncomplicated problem. So you know it has\na lot of steps and even for us it's one\nof the things that we have challenges\nall the time across the hardware\ncompanies. So while we doing this that\nyou know I'm basically kind of showing\nyou guys a little visualization to make\nit fun for me. I had a little you know\nuh pixel office that each of these\nagents basically you see are my agents\nthat do some sort of work. One of them\nis my benchmarker. One of them is my\nanalyzer. One of them is my uh skeptic.\nThat is really like hard rule that do\nnot accept what they pass to you. Go\ntest it yourself and make sure it's\ngood. Um you have your CAM bond that you\ncan always observe. So you can get a\nlittle bit fancy with this. Once you\nhave it ready for you this is just like\nyou know showcasing that for myself. you\nknow, I ask it to go run the\noptimization loop for this specific\nmodel for my AMD Mi325 GPU or 350 GPU\nand it basically goes through the loop\nof hey, I'm going to call the\nbenchmarker first. Then I'm going to go\nuh you know get the benchmark numbers\nand pass my analyzer. Analyzer found\nthat you know this is the good kernel to\nactually go and optimize and then later\non the analyzer passes that to another\none that is optimizer. the optimizer is\nsupposed to go write the code and make\nit optimized for you guys. And finally,\nthere's a skeptic that will be able to\nbasically um uh analyze the final\nresults and tell you if it's a good or\ngood result or not and passes back to\nme. So, this is this basically ran over\nI think maybe an hour or so. I just left\nit. I came back. I looked at the result.\nIt looked good. I showed it to my boss,\ngot a pat on the back and moved on. So,\nthis is like an example of what you\ncould do.\nGoing back to my uh to my agent, you\ncould see I applied the fix. Uh now I\nfound basically the the exact problems I\nhad. It fixed it and came back to me and\nshow me the results. So speaking of the\nmulti- aent part, I'll show you guys an\nexample of that as well. That'll be a\nwrap for the workshop part. I did have a\nchallenge. I don't know if you guys are\nrunning this with me. I will uh allow\nyou to fix to to see if you can, you\nknow, finish the challenge. But uh this\nis the the multi- aent part of it that I\nwant to make sure cover. So to add a new\nagent um the one like the one I showed\nyou guys have multiple agents. It's\nbasically as simple as open claw agents\nin the model and the agent name. So one\nof the most uh one of the most popular\nway of using openclaw is creating\nmorning brief. Say you know anybody has\ntheir own unique interests and you want\nto make sure that it's customized. Every\nmorning you wake up to something that\nyou like whether it's your financial\nnews, whether it's your GitHub repos\nthat you're watching. This is right now\nbe u I'm I've been told this is one of\nthe number one use cases of open claw.\nSo we're going to make that together\nright now. Um so the way you can\nactually add multi- aent is by basically\ndoing openclaw agents add and agent\nname. It will walk you through a very\nshort uh onboarding again. Do you want\nthe workspace to be here? Yes. Do you\nwant to copy the off? Yes. Uh configure\nmodel off for this agent now. No.\nAnd it should be able to basically start\nthe the agent bootstrap again for me to\nknow who it is and what we want to do.\nUh configure channels. No. All right. So\nthe agent is ready for me to use. And\nthe way I can use that agent within the\nsame agent I had now I have a main agent\nand I just added a second agent is again\nif I just do slash agent\nand I put the picker it's something like\nI don't know if you guys see it it has a\nlittle like name here. So main is my\ndefault now the morning brief. So I\nswitch to my morning brief and it says\nuh like hi you there.\nSo it should basically start again. It's\nlike a morning brief and now I give it a\ntask that um you want you want to\nbasically be the guy that always work\nwith me to have the best uh morning\nbeneath that I have because over time\nyou can actually go prompted to say like\nI don't like this or I want to make this\nmodification. So that's why you're\nactually creating a a second agent. Um,\ndo you guys maybe again if you guys know\nwhy I would not just have one agent to\ndo everything for me and just I need to\nactually have multiple agents because\nyou know it's connected to the same\nmodel. I'm not making a new model. The\nanswer was right. So it loses context.\nYes, because that's the whole nature of\nLLMs and he was right too. So there's\none down there. I guess uh the whole\nnature of LLM is attention based. So\nbased on that you have a limited\nattention to the to the model and the\ncontext length of it. So even if you\nhave a lot of instructions for one\nagent, you're not going to expect that\nto have the most powerful response for\nyou all the time because you know you\nhave given too many things. That's why\nyou want to separate them in in its own\nenvironment to make sure they have\naccess to their own tools and their own\npurpose. That's one of the reasons that\nyou want to maybe have a morning agent\ndedicated for that task.\nSo again, uh my name is Maddie. I'll\njust like basically do a very quick\nthing here. Uh you are morning brief\nagents. Um don't ask more questions\nand\nget ready.\nSo as simple as that this could start\nagain working for you and it will be\nable to\nuh create that morning brief that you\nwant for your you know whatever use\ncases here what I have for example is\nfor my own use case I want to know any\nchanges to SG lang project which is a\nGitHub project I'm not telling it where\nit lives it needs to go figure out how\nto get that information BLM project\nagain go figure out where it is find out\nthe information for me hugging face\ntransformer is and another project that\nI'm interested and at the end I say like\nyou know also like a summary of latest\nAI hardware news so go again read the\nnews yourself tell me what it is and it\nwill create like a version for me and um\nbased on that I can like prompt it again\nand say hey\num this isn't what I like this is what\nI'm not don't like and I will have again\nconfigurations to be able to maintain\nthat and and improve it over time so\nthis is going to go through that and do\nit uh probably respond to me in a\nminute. So, this is doing all of its\nresearch, doing all the reasoning. There\nis ways that it can actually prompt the\nopen uh claw agent to do um the verbose\nmode where it actually prints out a lot\nmore information for you that like what\nI'm doing, what I'm looking at. Uh the\nsimple way that when you set it up, it\nbasically goes the most basic uh\nimplementation of it, but you can again,\nas I said, you can um prompted to figure\nout what the problems with its own\nconfiguration is. That's a big hint for\nthe challenge part because the model we\nhave here coin 3.5 is a multimodel. It\ncan do images. It can do text and\nsometimes when you start the agent it\nonly configures as text right. So for me\nto start I was basically go say hey you\nhave image abilities go figure out your\nconfiguration and try to fix your image\nreading uh problems.\nWell this is still working. There's more\num commands that I can also tell you.\nOne of the most annoying thing is when\nyou, you know, have something that is on\na wrong track, you want to stop it.\nThere's keywords such as stop\nor abort that you basically can ask it\nto do. So, for example, a stop if I type\nit. Um\nmaybe I'll just do it for for the\npurpose of this while it's doing its uh\nthing just to show you guys. Or maybe\nshould I wait? I don't know. I'll show\nyou guys how to stop it. So if I do stop\nor say abort, it will basically take\nthat information tries to interrupt the\ncurrent process and um stop the running\nprocess at that moment. It is not very\ngood at that. I'll give you that in a\nbracket. So sometimes you have to do it\nlike multiple times to do it. That's one\nof the weaknesses of it and there's\nknown issue for it. But it's available\nthat works sometimes.",
  "transcript_chars": 31285,
  "ingested_at": "2026-05-21T19:16:49.647102+00:00",
  "source": "retry-no-transcript",
  "yt_meta": {
    "view_count": 374,
    "like_count": 9,
    "channel_id": "UCcIXc5mJsHVYTZR1maL5l9w"
  }
}