Every Level of Claude Agentic OS Explained

CChase AI
Computing/SoftwareSmall Business/StartupsManagementInternet Technology

Transcript

00:00:00if you don't know how to build your own agentic os then you are falling behind but not for the
00:00:04reason you think it's not because you need some fancy dashboard or a jarvis setup because that is
00:00:10not where the value of these systems lie the value is everything under the hood everything you can't
00:00:15see i'm talking about loop engineering skill architecture state management creating a second
00:00:20brain and being able to bundle it all together into a coherent customized product that works for
00:00:26you that is where the value lies and the skills required to build something like this can be
00:00:32applied to any project you work on with claude code which is why this is such a valuable thing to
00:00:37understand so in this video we are going to be breaking down the claude agentic os construct
00:00:42level by level so you can learn not only how to build one yourself but why it's so important that
00:00:48you do when we are looking at an agentic os like this one or this obsidian based one it's easy to get
00:00:54lost inside the visual spectacle of it all all these buttons and moving parts and metrics things
00:01:00you can't find inside the claude code terminal jump out at you right away and you kind of fall into one
00:01:05of two camps the first camp is like wow this looks really cool i want to get my hands on this i love
00:01:09all these like sort of visual things i can't find elsewhere and the other camp sees it for what it is
00:01:15which is just the visual interface and thinks this is just smoke and mirrors there's nothing here that's
00:01:21actually moving the needle forward and i think both camps miss something and what they miss are the ai
00:01:26fundamentals that are working under the hood that turn an ai os from just a fancy looking web app
00:01:32into a customized weapon you can use to attack any problem with claude code or frankly any model i'm
00:01:38going to talk about claude code today but this can be run again with something like codex or even a local
00:01:43model and when we talk about these fundamentals we can kind of break it down into four sections when it
00:01:48comes to this ai os the first level is the backbone and that is the skills in the loop engineering in
00:01:54the idea that we have codified everything we do inside of cloud code and we've turned it either
00:01:59into a skill or some sort of automation level two is memory and state control how can we make sure
00:02:06that our ai os has some sort of database of information that it can draw upon when we ask it questions and
00:02:12more importantly can we use this state this memory whether that's obsidian or something else you can
00:02:18use something like a standard database how can we use this in combination with the skills and automation
00:02:23we've built to create actually properly built loop engineered constructs that are somewhat self-improving
00:02:30if you watched my last video about loop engineering you kind of know what i'm talking about the idea that
00:02:34we need to be able to record things and set up loops that can see how we've done in past iterations to
00:02:40improve future runs and these two levels one and two is where we make most of our money this is
00:02:4590 of the value of any sort of ai os and once you have that locked in well then you can move into some
00:02:51of the cool visual stuff and that's sort of level three where we talk about the interface and the ui
00:02:56and sort of the customization aspect of this if you want to get outside of the terminal and also sort of
00:03:01spread your wings versus some of these desktop applications because the cloud desktop application
00:03:06is great but there's only so much you can do inside of it or set it up in certain ways and then lastly
00:03:10we have level four which is distribution because the cool thing about this is it doesn't have to be
00:03:14personal just to you you could share your ai os with members of your team or even clients and it's a
00:03:19great way to raise the floor in your organization a lot of what you do on level one and two can be turned into a
00:03:26literal button or voice command that you can implant inside of your ai os user interface that anyone can
00:03:34use they don't even need to run cloud code themselves it can all be done for them so those are the four
00:03:38levels we're going to be talking about today we're going to spend the vast majority of it here in
00:03:41levels one and two because the truth is you could still do all of this inside of your standard cloud
00:03:47code terminal or the codex cli or the codex desktop app if you really nail down these first two parts
00:03:52which is skills and loop engineering and memory and state this applies to any and all problems not
00:03:58just ai os but before we dive into level one a quick word from today's sponsor me so i just released my
00:04:04cloud code master class and it is the number one way to go from zero to ai dev especially if you don't
00:04:08come from a technical background we focus on real use cases it's updated every single week and it includes
00:04:14all of my personal builds so everything you see in today's video that i'm using for my demos you can
00:04:19find it here as well so if you want to get your hands on all that you can find it inside of chase ai plus
00:04:24there will be a link in the description so level one skill architecture loop engineering all this stuff
00:04:29what are we really talking about here well there's sort of four sub phases of level one we have the
00:04:35workflow audit we have the skill creation we have automation and then we have loop engineering and so step
00:04:41one is a workflow audit before we can create skills we need to know what you actually need to create skills
00:04:48for because remember why are skills arguably the most powerful thing in cloud code well because
00:04:53it allows us to get a specific output we're telling claude to do a specific thing in a specific way
00:04:58for a specific output but what sort of specific outputs do you need constantly throughout your day
00:05:04to day and your week to week seems like an obvious question yet most people cannot answer that and even
00:05:09if they can answer it they certainly haven't taken those workflows that they do manually inside of cloud code
00:05:15and turned it into skills or automated them so this is the first thing you have to do and this if you
00:05:20do nothing else will totally supercharge how you work with cloud code so here's a visual representation
00:05:25of what i'm talking about we have you and we have cloud code and for most people it stays just like
00:05:31this and it's a purely manual back and forth you open up the terminal you open up cloud desktop
00:05:36and you tell it to do certain things and inevitably you're telling it to do the same things all the time
00:05:41well what if instead we did an audit of everything you do day to day and week to week and all the ways
00:05:47you use cloud code and we codify that into skills you're doing the same tasks over and over
00:05:54why don't we actually consistent with the outputs and how they work now if you have used skills in any
00:05:58capacity the cell here is very obvious and i'm just telling you to do what you do with skills already
00:06:03but do that a hundred times more because undoubtedly the way you use cloud code whether it's as an
00:06:08individual or as a business can be broken up into a number of different domains for me that includes
00:06:14things like research content my online community my agency my sales on and on and on and underneath
00:06:20each of those domains are different things that i do specific tasks for content that includes stuff
00:06:27like hey i need to do outlines for all my projects i need to figure out the hooks for my videos i need to
00:06:32repurpose my content i need to create carousels on and on and on and on why are those not skills
00:06:38truthfully they should be and yet for most people they probably don't have a large robust set of
00:06:44skills like i do right here and this is the easiest way to improve cloud code now the practical question
00:06:49is how do you do this at scale and there are a few different ways we can do this number one is we do
00:06:56this purely via a manual process that means i go inside of cloud code i already have an idea of what
00:07:03it is i do i explain that task and then i turn it into can't type we turn it into a skill and obviously
00:07:12we're going to use the skill creator skill to do that right that's simple enough the issue with doing this
00:07:20especially if you haven't done it before is you probably haven't validated how cloud code should
00:07:25go about this the ideal scenario is you've done this manually you've confirmed it actually works
00:07:30and then you tell cloud code hey see how we just did that task now i want you to turn it into a skill
00:07:36and there's actually ways you can do that relatively quickly because remember cloud code has access to
00:07:42essentially all your previous sessions it can see the tool calls it can see what you told it to do it
00:07:46has the full back and forth we can really get a head start on creating this sort of like skill repository
00:07:52by telling cloud code hey i want you to take a look at our last three sessions five sessions 10 sessions
00:07:5920 sessions i want you to pull out everything we've done and give me a list of things that we can turn
00:08:05into skills that i do all the time so that becomes option number two which is we have it look at previous
00:08:11sessions and sort of pull the work out of us this way it's actually going off real data you know this
00:08:17isn't a guessing game of what you think you should do in cloud code it's actually going to take a look
00:08:21at what you've done and so that prompt can just sound something like this hey can you go through our
00:08:28last 10 sessions that i've had back and forth with you and i want you to sort of pull out some repeated
00:08:34tasks or things we've done over and over again that aren't skills yet but that i want to turn into
00:08:39skills and create some sort of chart showing what that task is what the output should be and the
00:08:44proposed skill so that's pretty much it doesn't have to be fancy you can talk to it in plain language
00:08:49and like you can see right here what is it doing it's going to find our session files first now there's
00:08:55actually a third option for how we can approach this option number three is we have it do an interview
00:09:03right we have clawed code interview us and we say hey i'm just going to give you a stream of
00:09:09consciousness about what i do day to day week to week and i want you to ask any questions if there's
00:09:13any blind spots and then i want clawed code to pull out of that conversation these sort of tasks that can
00:09:20be turned into skills so same sort of idea it'd be a same sort of simple prompt let's pull it up right now
00:09:27and then i would say something like i'm trying to turn all my daily and weekly tasks into skills when
00:09:36they make sense so i'm going to start off by giving you a stream of consciousness of sort of like what i
00:09:41do each day and then i want you to sort of turn it into an interview and call out any blind spots because
00:09:47at the end of this i want you to have as much context as possible about what i do and sort of the
00:09:51outcomes i'm looking for and i want you to pull out specific tasks and i want to be able to turn these
00:09:57tasks into skills and eventually automations
00:10:02that's pretty much it doesn't have to be fancier than that the whole idea is that we are just
00:10:07throwing as much context as possible about our work about our day about our week into cloud code
00:10:14and codifying it we kind of want to just turn what we do into a checklist this is kind of the
00:10:18saying the way you should approach this is like you brought in some person to be your personal
00:10:22assistant you want to offload as many tasks onto them as possible well how would you do that
00:10:27well it's pretty obvious you would tell them what you do and then give them step-by-step instructions
00:10:32to do that that's all we're doing we're just doing it with cloud code and yet most people don't do this
00:10:39and because these are skills these are now tangible workflows that we can look at and edit as needed
00:10:43until we really dial in those outputs and right here you can see one of the repeated tasks that
00:10:48pulled out from previous sessions things like checking for tool and repo updates for videos
00:10:53and it continues to go on and on with these tasks it found now creating these skills is just layer
00:10:59number one of this right we've codified it and oftentimes these are things that are going to get
00:11:04repeated over and over well if they're going to get repeated over and over is there any reason why we
00:11:08don't simply turn these into automations right again we want to move away from this manual approach to
00:11:14everything well like add this task i did manually now it's a skill it's been codified well now let's
00:11:20just set it up into an automation if it makes sense and this is so easy to do inside of cloud code we can
00:11:25quite literally just prompt it can we turn that skill into an automation and if you want a more visual
00:11:30approach you can set this up very easily inside of cloud desktop we just go to routines we give it a
00:11:36name so that'd just be like auto one the instructions would just be run this skill insert the name of the
00:11:43skill and then we would set it on a specific schedule that's pretty much it and the third level here would
00:11:49be setting up some level of loop engineering now i'm not going to go ultra deep into loop engineering
00:11:55because the video i put out yesterday certainly does but we have sort of the foundation for creating
00:12:00strong loops we have the skill we've turned it into an automation now it's simply a question of well
00:12:06are we taking a certain automation and are we trying to add some sort of like self-improvement loop
00:12:12to it and this will tie into memory and state as well but that's all i'm really going to mention
00:12:17there just understand if you're someone who's trying to dive into loop engineering and that whole
00:12:21side of the equation you're set up very well to do this given we laid the foundation with our skills
00:12:27and our automations but zooming out here for a second this is the backbone of everything it's
00:12:34codifying your life and making it so we can get consistent outputs from clawed code for the tasks we
00:12:40actually care about and as you can imagine that kind of has nothing to do with these fancy dashboards
00:12:46and all this other stuff that comes along within agentic os i love the fancy dashboards i think they're
00:12:50really cool but this is the power and you can imagine applying that to any problem and any sort
00:12:56of use case with cloud code is is very simple and straightforward it doesn't require any of the
00:13:00other levels which is why i think it's important to sort of be able to create your own aios because
00:13:05as we stack these levels on top of one another you can see how sort of modular and flexible this is so
00:13:12workflow audit talked about the three different ways to do that right we can do it manually we can have
00:13:15it look at our previous sessions or we can have it run an interview once we've done that we have it
00:13:19create the skills then we ask ourselves hmm can any of these skills be automated and then lastly we kind
00:13:25of have that discussion about does loop engineering make sense for this particular use case and speaking
00:13:29of loop engineering that brings us into things like state and memory which is level two now a lot of what
00:13:37i'm going to talk about today will be in the context of obsidian but understand it doesn't have to be
00:13:42obsidian everyone likes talking about obsidian because it's free and it's relatively easy to understand
00:13:47everything that works with obsidian can also be done inside of a traditional database
00:13:52right it doesn't have to be obsidian obsidian is just simple to use and really even more so than
00:13:57obsidian it's the idea of file structures and setting up cloud code in a coherent manner truthfully you
00:14:05could probably have no database and you could have no obsidian and if you just set up cloud code with a
00:14:11file structure that is coherent and makes sense you're like 99 of the way there obsidian just makes
00:14:17it easy and databases obviously have their own power that goes along with them so how should we set up
00:14:22this part of the equation well we're going to answer that question through the lens of obsidian and what
00:14:27the file structure should be doing for us now obsidian like i mentioned before is completely free the idea
00:14:32with obsidian is that you simply download it and then you're going to designate a folder on your
00:14:38computer as the vault it can be any folder or it can be a brand new folder now when you're installing
00:14:43obsidian you'll see a pop-up like this where again you can create a new vault or you can open a folder as
00:14:49a vault now when we're talking about a vault again it's literally just a file so you need to ask yourself
00:14:54for this agentic os what file should it live in what file is going to have all the information i
00:14:59wanted to know about so if we're treating it like a personal assistant perhaps one of the domains you're
00:15:05going to have it help you out with is like sales so you have a lot of sales data well whatever file
00:15:09i designate as the vault i'm going to put want to put at least a copy of all my sales data in there
00:15:14that's sort of the idea now once we've designated a folder as the vault in order to connect cloud code
00:15:19to obsidian it's as simple as just opening cloud code inside obsidian
00:15:23so i navigate inside my terminal to whatever i called that folder in this case i literally called
00:15:29my vault the vault so i'm now inside the vault next i just open up cloud code
00:15:37and boom cloud code for all intents and purposes is now connected to my vault so i now have my vault
00:15:44and cloud code is open and connected to it now the question becomes how do i actually set up my file
00:15:51structure inside the vault this is not a trivial question this is very important because the whole
00:15:56value add for something like obsidian is the idea that cloud code can be connected and open inside of
00:16:03some folder that has a bajillion subfolders and a bajillion files inside of those folders and you the
00:16:11human being can ask cloud code a question about any of the files and folders inside of here and give you
00:16:17an accurate and quick answer how do we do that well it's going to all come down to how we sort of set
00:16:23this up right if we just have one folder with 10 million files in it and there's like no backlinks
00:16:29and nothing is connected and there's no sort of hierarchy claude's going to struggle to quickly
00:16:34find you an answer and when i mean it's slow i'm also meaning it's going to use more tokens and
00:16:39ultimately it's going to cost you more money so we have to set this up in a clear manner your mental
00:16:44model here is setting up a map for cloud code right we're looking at the knowledge graph of my obsidian
00:16:49vault these are all the files inside my vault and how they connect to one another the ideal scenario
00:16:56is that when i tell cloud code or ask it a question about something inside this giant morass of files
00:17:01it has a very clear path to finding that file and therefore finding my answer so think of obsidian
00:17:08as essentially your filing cabinet for everything now there's become somewhat of a common way to set
00:17:15up your file structure inside of obsidian when we're talking about cloud code and this comes from carpathy
00:17:21this tweet from carpathy got over 20 million views and it was all about how he sets up his knowledge
00:17:27base using obsidian so that large language models like claude can access information quickly and
00:17:33effectively and it goes something like this we have the primary vault which you are now familiar with
00:17:39and then we sort of have three subfolders we have one subfolder which we call the slash raw
00:17:47folder ra w this folder is where all unstructured data goes beneath that we have another subfolder
00:17:56called the wiki folder the wiki folder is where we have taken the unstructured data think hey we just
00:18:03did all this research about i don't know ai agents write a bunch of unstructured data includes a bunch of
00:18:10articles and all that sort of stuff and we've turned it into structured data we've now turned that
00:18:15into like a wikipedia style article about ai agents so the data and like what's going on here and
00:18:22everything we want to know about it is very clear right this is instead of trying to go through you
00:18:26know 20 different like source documents we now have everything nice and neat under this wiki file
00:18:32the third subfolder in the vault is essentially for outputs so let's say i did my research about ai
00:18:40agents which is now in the raw folder folder number one i then turn that into structured data into a
00:18:45wikipedia article which is in folder number two now i want to turn that into say and let me move this
00:18:51over here let's say i now want to turn that into a i don't know a slide deck okay some sort of clear
00:18:57deliverable right we're not just talking about data here we've actually turned it into something useful
00:19:01well that would be folder three okay and so the idea is for most of the data we play around with
00:19:07we can sort of set it up like this unstructured structured and then outputs that is the carpathy
00:19:15obsidian rag and i say rag and air quotes because this isn't a true rag system but that's one way to
00:19:20do it now the beauty of this carpathy system isn't necessarily that we split it up into unstructured
00:19:27structured and outputs the real beauty is that at every level of this we have an index dot md file
00:19:36right this is just a text document a markdown file that is telling clod code at every level we go down
00:19:43what it's looking at so for example if i'm talking to clod code i'm talking to my aios and i say hey
00:19:49i want you to give me all the information about ai agents remember we created a wiki article about ai
00:19:56agents well first thing it's going to do is it's going to look in the vault and it's going to hit
00:20:01this index.md and this index.md is going to say hey inside at this level of our system we have a raw
00:20:08file for unstructured data a wiki for structured data and outputs for things like slide decks and
00:20:12that sort of thing so clod code knows right away hey he wanted information about ai agents so we're
00:20:18going to go to this wiki article now inside of the wiki article inside of the wiki folder rather guess
00:20:23what's inside here there is also an index.md now we have one article in here so does it really need
00:20:31essentially a table of contents for a folder with one thing in it no of course not what if you've been
00:20:37using this for a year or two or five and you have not one or two or ten but you have thousands of
00:20:42documents inside of here and potentially subfolders as well well an index.md is going to make it way
00:20:47easier for cloud code to hit that folder and understand what it's looking at and where it
00:20:51needs to go because remember what is the purpose of all this it's to give cloud code a map and if it
00:20:57every single new room it enters there's a clear spot it can go to and figure out what it's looking at it's
00:21:03going to be faster and it's going to be cheaper and it's important to understand that the power comes
00:21:08from that not necessarily the somewhat arbitrary folders we created it just needs to know where it's going
00:21:14you don't have to do raw you don't have to do outputs you don't have to do any of this carpathy
00:21:18stuff you just need a map for cloud code that makes sense and it's probably going to be unique
00:21:22to you because your structures of your data and what you're trying to do are always going to be unique
00:21:28so you have a better answer than anyone for how you should structure this now if you don't have
00:21:32that answer guess who can help you cloud code simply tell to look at your vault and say hey
00:21:36what sort of structure makes sense oh use carpathy's obsidian rag setup for inspiration that will kind
00:21:43of do everything you need it to do the only other thing i would mention would be hey let's create a
00:21:48claw.md file that talks about this in my claw.md file inside of my vault it talks about my vault
00:21:54convention specifically the vault structure aka what sort of files and folders it's looking at you can see
00:21:59over here on the left i don't just have three i have several i have content notes runs inbox ops
00:22:04projects etc etc furthermore i have a whole thing about the navigation pattern essentially saying hey
00:22:09when you're trying to find something here's the path you should follow and i think using that sort
00:22:13of template is very flexible you can apply it to any sort of structure or any sort of data you're working
00:22:18with and you can come up with something that will be effective for you so when we talk about level two
00:22:24that's what we're talking about we're giving cloud code a map and it needs to make sense now this also
00:22:29plays into you know loop engineering because and skills and automations because all these outputs
00:22:34need to go somewhere and they need to be logged and they should be logged in a way which you know what
00:22:38i'm going to say makes sense for loop engineering specifically when we talk about self-improving
00:22:44skills and automations well for that to work we need somewhere where cloud code can see what the past runs
00:22:52have done or really what the loop should be able to see what past runs it's done so then it can make
00:22:56future improvements and this should all be tied together in the same place and if you master
00:23:00those two levels you have 90 of the power of an aios already at your fingertips and all this can be
00:23:08done through the terminal or the desktop app or anything because this point you've kind of codified
00:23:12your workflows and now you have a way to actually see what's going on record what's going on sort of
00:23:18create your second brain and allow cloud code to pull out insights you otherwise wouldn't have
00:23:23in an efficient manner now level three is where we put a custom visual wrap around everything we've
00:23:28done up to this point and we have a few options with how we do this obviously it can be purely
00:23:33customized but we can make this something that is purely web app based or something that is obsidian
00:23:37based now when we talk about web app based we're talking about something like this again this is
00:23:43cloud code under the hood it's connected with obsidian but i now have a bunch of custom metrics
00:23:48that i have tuned for what i need to see so for me on the left it shows stuff related to my content
00:23:54creation right what's my youtube subscriber account instagram my latest video my clawed five hour window
00:23:59i have directives that are pulled from my google calendar i can look at documents it's created over
00:24:03here on the right i've taken a bunch of my automations and skills and i've turned them into just
00:24:09single buttons so if i just click on something like inbox brief you can now see that it's queued it's
00:24:15popped up right over here and under the hood claude is running going through my inbox creating drafts
00:24:21and it's going to let me know what it thinks is important and the cool thing about this is again
00:24:25you can change whatever metrics are shown here to be whatever you want the whole idea of this being
00:24:31a floor raising mechanism the idea being i can now give some of the power to cloud code to non-technical
00:24:37team members and clients we'll talk about that in level four has a lot to do with what you see over
00:24:41here on the right where we've turned automations and skills into a button you can literally press
00:24:46so instead of telling them hey learn how to use cloud code here's a skill you need to install here's
00:24:50how you run the skill here's how you automate it now i'm just going to go ahead and i'm going to set
00:24:55up this aios for them and now they can just click a button and it will do all that for them and will
00:25:00either dump it inside their own obsidian or a team's obsidian inbox brief is done inbox brief ready
00:25:0832 threads triaged with a gear up contract and the open ai merge campaign flagged urgent
00:25:15all right that's enough out of you but yeah you can hear in this case i also have a voice
00:25:19model attached to it so i can talk to it it can talk back and that voice model is completely local
00:25:24by the way right that's running on my actual computer that isn't going out to 11 labs so it's free
00:25:29if i click on the inbox brief it brings up the entire write up and this too is something
00:25:36that i can open inside of obsidian and speaking of obsidian you can also create in a command center
00:25:41a visual layer inside of obsidian itself and that's what we see right here so the metrics are a little
00:25:47bit different but they're similar i can see my token burn i have you know the same sort of setup where
00:25:53i click a button and it runs skills or automations i have different tabs so for me i want more insight
00:25:58into sort of like audience metrics that's going on the content side as well as some research stuff so
00:26:04you can make this extremely customized and again the cell here at this point isn't the fancy visuals
00:26:10it's that i can have a one-stop shop for a lot of different things that are a little harder to have
00:26:15visibility into when i'm strictly inside the terminal now in terms of how you would create something like
00:26:21this the web app is just like creating any sort of web app using cloud code you're going to give
00:26:26it some sort of visual idea i mean obviously my exact setups you can find inside of chase ai plus but
00:26:32you should just find some sort of you know website or setup you like you can take a screenshot of this
00:26:37you would dump it into cloud code and you essentially say like hey here's all the skills i already use
00:26:42i want this connected to the vault here are the metrics that are important to me that i want to see
00:26:47in one place let's go ahead and create this let's put a visual wrapper over all this and the same
00:26:52deal goes for obsidian obsidian runs in a plug-in system so you're pretty much creating an app but
00:26:58it's specifically for obsidian and if you just tell claude code to say hey can you sort of take
00:27:05the web app we just created and create an obsidian plugin version of that again it will give you something
00:27:10like this and you just install it and run it from obsidian now a quick note about what's going on
00:27:15under the hood here now under the hood if i click one of these buttons which again are related to
00:27:20skills let's say i click the morning brief skill what's actually occurring well this is essentially
00:27:24calling on a headless version of claude code so it's just like as if i opened up my terminal
00:27:30and i have a version of claude code running and it's now going to run you know forward slash morning
00:27:37brief the difference is when i click that button it's doing a headless version of it so this terminal
00:27:42doesn't literally pop up on your computer it's headless it's invisible and it uses a command called
00:27:49claude dash p now there was some drama with claude dash dash p not too long ago because anthropic came
00:27:55out and said hey if you use claude dash p it's not going to pull from your claude mac subscription it's
00:28:01going to pull from this 200 credit which is tied to api costs that was kind of a problem although they've
00:28:07sort of walked back from that and that isn't something that has occurred yet so for now this is
00:28:12still pulling from your max plan so it's the same as if you open up your terminal and ran it and that's
00:28:16how we're able to create these sort of structures that still call on claude code we get all the power
00:28:20out of claude code but it's done invisibly behind the scenes so when we talk about level three what is
00:28:25this bias well by this customization we aren't locked into the terminal or the desktop app we can have it
00:28:30show whatever we want we also have the ability to give this to members of our team this is what we talk
00:28:36about in level four which is distribution i mentioned a little bit earlier if i hand someone
00:28:41this web app and it's tied to all these different skills they're just one click away from getting a
00:28:45lot of power out of cloud code because remember all the power comes from those skills and automations
00:28:49if i make it super easy for someone to use those well it's kind of like spinning them up on claude code
00:28:54without actually spinning them up on cloud code the obvious question then becomes well how would you
00:28:58actually distribute to them and there's a few options when we're talking about something that's
00:29:02web-based it's actually much easier so giving something this somebody this versus giving them
00:29:07the obsidian version is much simpler because since it's web-based i can put it on github i can you
00:29:12know create a whole zip folder it's very easy for me to transfer it to them and have it get get it up and
00:29:17running versus something like obsidian obsidian is a little harder to work that way so obsidian would
00:29:23require a little more hands-on work from you so if you're like hey i really like the sort of like
00:29:27obsidian command center deal how would i bring that to a member of a team well you would kind
00:29:31of have to set it up for them it's not as straightforward it's not much more difficult
00:29:35but it isn't as simple as like hey there's a github repo go ahead and clone it and point cloud code at
00:29:40it but again the customization piece is a huge sell here especially for those of you who are doing
00:29:45any sort of client work i can't tell you the amount of people who want to use a and want to use claude
00:29:51but are totally turned off and frankly just really scared of the terminal and even the desktop app like
00:29:56it is a bridge too far for most people if you're watching this video you probably scoff at that but
00:30:00i'm telling you you live in a bubble that 99 of people just like won't go there no matter what
00:30:05you do and being able to say hey instead i'm just going to throw you this and i'm going to set it up
00:30:09for you and you just either talk to it with a voice or you press a few buttons it goes a long way
00:30:14this sort of dashboard effect with a non-technical population like genuinely needs to be studied
00:30:19because it changes how people interpret these like technical tools and so zooming out now that we've gone
00:30:24over levels three and four you can see they're really just the cherry on top of the ai os almost
00:30:30all the power almost all of your time should be invested in these first two levels the skills the
00:30:37loop engineering is the automation the codification the memory and the states right can you do the same
00:30:42things with cloud code every time can you be consistent and can we log those things and essentially
00:30:47create the second brain for cloud code that it can not only reference but use to improve upon what it
00:30:52already does if you can figure out that then you're going to be well ahead of the general population
00:30:59when it comes to using this tool so that is where i'm going to leave you for today's video i hope
00:31:03breaking down ai os and sort of these four levels made it a bit more clear about how these things work
00:31:08where the value is and how you can create something like this yourself like i mentioned earlier if you
00:31:12want my exact setups that's all inside of chase ai plus but other than that let me know what you thought
00:31:18and I'll see you around.

Key Takeaway

An effective agentic operating system functions by codifying manual workflows into a repository of skills, automations, and a structured knowledge base, which then allows for consistent, self-improving outputs.

Highlights

  • Building an agentic operating system centers on skill architecture, loop engineering, and state management rather than visual interface design.

  • Ninety percent of the value in an agentic OS comes from the backbone of codified skills and automated routines, not the GUI.

  • Workflow audits require identifying repeated tasks across domains like content creation and sales to convert them into reproducible skills.

  • Memory and state control rely on structured file systems, such as obsidian vaults, to provide AI agents with a coherent knowledge map.

  • Headless execution of Claude Code allows for automated task processing in the background, independent of terminal visibility.

  • Distributing an agentic OS to non-technical users involves wrapping workflows in simple, button-driven interfaces to bridge the gap between complex terminal tools and end-user accessibility.

Timeline

The Four Levels of Agentic OS

  • Agentic OS value resides under the hood, specifically in skill architecture, loop engineering, and memory management.
  • Visual interfaces are secondary to the underlying AI fundamentals that enable customized problem-solving.
  • The OS structure is divided into four levels: backbone (skills), memory/state, interface/customization, and distribution.

True agentic OS utility does not come from fancy dashboards but from robust engineering of internal processes. The backbone consists of codified skills and automation, while level two focuses on database and state control. Levels three and four provide the interface and distribution capabilities for team collaboration.

Level One: Backbone and Workflow Audit

  • A workflow audit identifies specific, repeated tasks to convert into codified skills and automations.
  • Skill creation methods include manual definition, analyzing previous session logs, or interviewing the AI to identify blind spots.
  • Loop engineering adds a layer of self-improvement where the system records past iterations to enhance future performance.

Effective workflow automation begins by identifying day-to-day tasks that are currently performed manually. By auditing these tasks—either manually, via session analysis, or through an AI interview process—these workflows become persistent, actionable skills. Once codified, these skills form the foundation for automations and self-improving loops.

Level Two: Memory and State Control

  • File structures, such as Obsidian vaults, act as the database for an agentic OS, allowing for context-rich interactions.
  • An index.md file at every folder level provides the AI with a map, reducing token usage and improving query speed.
  • A logical file structure using raw data, structured wiki-style knowledge, and clear outputs is essential for effective AI navigation.

Establishing a coherent file structure is critical for enabling an AI agent to retrieve and use information efficiently. By adopting a convention—such as the one popularized by Andrej Karpathy—files are organized into raw data, structured wiki documents, and finalized outputs. The inclusion of index.md files ensures the AI can quickly navigate this knowledge base.

Level Three and Four: Interface and Distribution

  • Custom interfaces, whether web-based or within Obsidian, turn complex automation chains into simple, clickable buttons.
  • Headless execution of Claude Code enables invisible, automated background processing without requiring manual terminal interaction.
  • Distributing an agentic OS raises the capability floor of entire organizations by providing non-technical users access to sophisticated automated workflows.

Custom visual layers wrap the underlying logic, allowing for the trigger of complex workflows through simple user interface elements. These interfaces make powerful automated tools accessible to non-technical users, significantly lowering the barrier to adoption. This approach enables the scaling of sophisticated AI capabilities across teams and clients.

Community Posts

View all posts