The argument in one line.
Building an AI operating system is fundamentally a data organization problem, not an AI problem—and solving it requires hygienic file structures, global vs. local skills architecture, and a database layer before layering on any interface.
Read if. Skip if.
READ IF YOU ARE…
- A founder or solo operator running a service business who manages multiple ongoing projects and wants to visualize agent activity and task completion across your entire operation.
- A developer or technical operator already comfortable with Claude API, SQLite, and Telegram who wants to see how multi-agent systems can be architected around a shared memory layer.
- Someone building or scaling an AI-powered workflow who sees your current setup as brittle and wants to understand how to structure agents, memory, and task delegation as a data problem first.
SKIP IF…
- You're looking for a no-code, off-the-shelf tool you can spin up today — this is a custom-built system that took hundreds of hours to develop and requires engineering work to implement.
- You're a non-technical creator or operator without database, API, or backend experience — the architecture assumes you can read code and modify Claude prompts at minimum.
- You need step-by-step instructions to replicate this exact setup — the video is explicitly a concept tour, not a build guide or technical tutorial.
The full version, fast.
An AI operating system is a data engineering problem dressed up as an AI problem, and the bottleneck is how cleanly your files, skills, and agent memory are organized � not which model you run. The mechanism is a hive mind of specialized Claude Code agents that share one local SQLite database, expose themselves through Telegram, and inherit every globally installed skill, CLI, and integration; a cheap auxiliary model auto-routes incoming tasks, cron jobs schedule recurring reports, and a war room runs slash-commands like stand-up and discuss against the shared memory. Build the boring list-view and database first, layer visualizations, scheduling, and voice on top later, and treat the whole system as iterative rather than a one-shot configuration you finish.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:34
01 · Cold open — Hive Mind 3D
Opens on the 3D hive-mind brain showing every agent's completed tasks as glowing nodes. Filters by agent (main / meta / comms / content / ops / research), 205 entries shown.
00:34 – 00:59
02 · 2D filter + list view
Obsidian-style 2D graph, then drops to a raw list view of every agent action, action type, and summary — the source of truth that powers the prettier views.
00:59 – 02:27
03 · The War Room — slash standup
Opens the 'war room' chat. Runs `/standup` and every agent (research, comms, content, ops, main) chimes in with what they did in the last 24 hours; main agent synthesizes the team summary.
02:27 – 03:24
04 · What this video is and is not
Sets expectations: not a step-by-step zero-to-hive-mind tutorial — this took hundreds of hours. The goal is to demystify 'AI operating system' as a data-organization exercise with layers on top.
03:24 – 04:24
05 · Lineage — from OpenClaw to ClaudeClaw V3
Whiteboard: 'The Wrapper' diagram. Telegram is the steering wheel, the Anthropic SDK is the bridge, Claude Code on your local computer is the engine. Evolved from a single agent → team → war room → full hive mind.
04:24 – 06:39
06 · Meta Ads agent — concrete example
Hand-drawn 'Meta Agent Example' diagram: drop in the meta-ads-cli globally → every agent inherits it → a scheduled task fires every morning at 7:30am with ad performance, blind spots, hot-takes, and direct hyperlinks to each ad.
06:39 – 07:13
07 · Mission Control dashboard
Glorified Kanban with columns per agent. Create a task, drag to an agent, or hit auto-assign and let Gemini Flash pick the best-fit agent (cheap classifier model). Tasks queue → run → done.
07:13 – 08:00
08 · Auto-assign with Gemini Flash
Sample task ('Send Mark an email saying hi') auto-routes to the comms agent. Gemini 3 Flash runs a dynamic prompt seeded with every agent's description and classifies — costs are inconsequential. Telegram is swappable for Slack/Discord.
08:00 – 09:12
09 · Sponsor — Early AI-dopters / Claude Code Magic Course
Soft pitch for the first link in description: skool community, the carbon-copy ClaudeClaw kit, coaching, plus the Claude Code Magic Course (Zero-to-Hero, Command Center, ecosystem lessons).
09:12 – 09:25
10 · Comms drafts the email — behind the curtain
Returning from sponsor: comms uses the Google Workspace CLI (already in Claude Code) to draft the Gmail. Shows the Gmail draft on screen. Stresses 'symbiosis' between front end and back end APIs.
09:25 – 09:42
11 · Mission Control wiring
Whiteboard: Gemini classifies → queued → running → done. The hardest part is keeping the dashboard perfectly synced with Telegram and the various API integrations.
09:42 – 10:30
12 · Scheduler tab — cron under the hood
Whiteboard: 'What the system sees: raw cron. What YOU see: every morning at 6 AM, weekdays at 6:30 PM, Sundays at 9 AM.' A translation layer renders cron into plain English.
10:30 – 12:00
13 · Agents tab — command center for the team
Per-agent cards: Main (Opus 4.6), Meta, Comms, Content, Ops, Research (Sonnet 4.6). Swap models, edit personality, stop/restart, plus a suggestions feature that uses Gemini Flash to scan conversations and flag overloaded agents (Comms is doing too much — maybe spin out an email manager).
12:00 – 13:00
14 · Creating a new agent
Click 'new agent', name it, give it a display name and description, paste the Telegram bot token — and underneath each agent is just a CLAUDE.md + a YAML config. Minimalistic by default; optionally layer agent-specific skills and rules.
13:00 – 13:32
15 · Unified Chat tab
Web-based chat with every agent in one place. Same harness, same Claude Code subscription, but consolidates all your agent conversations off Telegram and into a single pane.
13:32 – 15:56
16 · Memory systems — salience, recency, importance
Five to six memory layers rolling up into three categories (importance, salience, recency). Searchable 'blurred memories' tab. A `/insights` cheap-model pass derives meta-insights about how you've used the system over the last 30 days. Local SQLite for everything — no cloud DB required.
15:56 – 17:44
17 · Hive Mind 3D — the holy grail view
Reiterates: the list view is the foundation, the 3D and 2D views are pure visual layers on top. Mark recreated the Obsidian graph experience by Loom-recording his wishlist, feeding the video into the Gemini skill's video-understanding API, and having Claude Code build the spec.
17:44 – 20:09
18 · War Room — voice + text meetings
Voice mode (with a launch jingle) lets him talk synchronously to all agents. Text meeting mode adds room-specific slash commands: `/standup`, `/discuss`, plus the ability to pin one agent as the meeting lead. @-tag any agent to direct a message, custom-GPT style.
20:09 – 23:04
19 · The AI OS paradigm — your back of house
Big idea: the dashboard layer is cute, the foundation is everything. If your desktop is chaos, layering ClaudeClaw on top won't save you. Decide what becomes a global skill vs. a project skill, fix your file hygiene, then add memory, scheduling, and a remote-control surface (Telegram / Signal / Discord). 'This is a data engineering problem, not an AI problem.'
23:04 – 24:01
20 · Resources + CTA
Promises the system blueprint as a feed-to-Claude-Code reverse-engineering kit, plus a way to drop the entire video into Gemini and have it generate the perfect prompt + supporting docs. CTA back to the Early AI-dopters skool community in the first link.
24:01 – 24:12
21 · Outro
Standard like / comment / see-you-next-time outro.
Lines worth screenshotting.
- A hive mind visualization where every node represents a completed agent task makes the activity of an entire agent workforce scannable in seconds rather than requiring log reviews.
- Building an AI command center is fundamentally a data organization exercise with intelligence layers on top — treating it as an AI problem first is why most implementations fail.
- When Claude Code is already connected to external services via CLIs and integrations, every new agent inherits that infrastructure automatically without additional setup.
- A Meta ads skill that pulls campaign ROAS, spend, and blind spots into a daily Telegram report at 7:30 AM turns a novice advertiser into someone who can act on data every morning.
- A slash standup command that invokes every agent simultaneously and produces a synchronized status report replaces what used to require a team meeting.
- Skills shared globally across all agents in a hive mind mean a new agent starts with the full capability set of the existing workforce from its first task.
- Connecting Claude Code to Telegram via the Anthropic SDK means every agent in the workforce can reach you on mobile and every conversation can reach back into the agent system.
Steal the architecture.
ClaudeClaw playbook for Joe
Build the list view first, ship the brain second, and never demo the chrome until the data layer is rock solid.
- Lead every Mod-anything walkthrough with the artifact, not the promise. Mark's cold open is a glowing 3D brain, not a sentence. Joe's Mod Producer / Paperclip demos should open on the most photogenic running screen, with the talking head as a corner bug — not centered.
- Stop building agents in a vacuum — build a 'list view of everything every agent has ever done' first, then layer 2D/3D visualizations on top. That same 'log table' pattern would let Joe ship a hive-mind for JACE/REESE/SAGE/RYDER inside a week without a single fancy renderer.
- Use the cheap-classifier pattern. Gemini Flash auto-assigns the task, Claude Code does the work. Joe should bake this into Paperclip orchestration today — it cuts token spend on routing decisions to near zero.
- Whiteboard your architecture as a metaphor before you build it. Telegram = steering wheel, Claude Code = engine. One sentence does what 30 minutes of explanation can't. Joe should do this for the $6 Stack and the MCN+ membership.
- The CTA placement is a master class — mid-roll soft, end-roll harder, both anchored to a single Skool link in description position #1. No multi-link confusion, no 'check the description for all my links.' One destination.
- End with the thesis, not the features. 'This is a data engineering problem, not an AI problem' is the line that travels. Joe's 'Stop renting / Own your stack' has the same shape — use it as a closer, not an opener.
Terms worth knowing.
- hive mind (AI agents)
- A shared memory and state layer connecting multiple specialized AI agents — allowing them to see each other's completed tasks, coordinate work, and collectively give the operator a unified view of all agent activity in one place.
- AI operating system (AIOS)
- A personal or business-level harness that organizes an AI model's skills, memory, integrations, and scheduling into a coherent operating environment — analogous to a computer OS that provides structure for applications running on top of a CPU.
- hive mind graph view
- A visual representation of all agent tasks and knowledge connections displayed as a network of linked nodes — inspired by Obsidian's graph view — used to see patterns, activity volume, and relationships across an agent workforce at a glance.
- CLAUDE.md / YAML agent config
- The two files that define an AI agent's identity and behavior: a CLAUDE.md markdown file containing the agent's persona, rules, and skill references, and a YAML configuration file specifying its technical settings, model, and resource access.
- cron job
- A time-based task scheduling mechanism on Unix-like and server operating systems — defined by a string of numbers and symbols specifying when a command should run automatically, such as every weekday at 07:30 AM, without manual triggering.
- SQLite
- A lightweight, file-based relational database that runs locally on a computer without a separate server process — used here to store all agent conversations, tasks, memories, and hive-mind data in a single free local file.
- vector database
- A database optimized for storing and searching data as high-dimensional numerical vectors — used in AI memory systems to find semantically similar past memories or documents based on meaning rather than exact keyword matches.
- salience (memory systems)
- The relative importance or prominence assigned to a memory — used in AI agent memory architecture to determine which memories should be retained long-term, surfaced frequently, or allowed to fade as they become less relevant.
- embeddings
- Numerical vector representations of text produced by a language model — used in semantic search and memory systems to measure how conceptually similar two pieces of text are, enabling an agent to retrieve relevant memories based on meaning rather than exact words.
- Anthropic SDK
- The official software development kit for Anthropic's Claude models — providing developers with a programmatic interface to send messages to Claude, manage conversations, and integrate Claude's capabilities into custom applications and agent systems.
- Telegram bot token
- A unique credential issued by Telegram's BotFather that authenticates an AI agent as a specific Telegram bot — used to connect a Claude Code agent to Telegram so it can receive instructions and send responses via the messaging app.
- ROAS (return on ad spend)
- A metric measuring how much revenue is generated for every dollar spent on advertising — calculated as revenue divided by ad spend — used to evaluate the efficiency and profitability of paid advertising campaigns.
- Meta CLI (ads)
- A Meta Platforms command-line interface that allows programmatic access to Meta advertising account data — enabling AI agents to query campaign performance, ad spend, and audience metrics without navigating the Ads Manager web interface.
- Gemini Flash (model)
- Google's lightweight, low-cost language model in the Gemini family — used here as an inexpensive workhorse for classification, summarization, and routing tasks where high intelligence is not required but large context windows and low cost per token matter.
Things they pointed at.
00:06productHive Mind (ClaudeClaw V3 dashboard)
03:50toolAnthropic SDK
04:30toolMeta Ads CLI (meta-ads-cli)
05:50toolGemini skill (with Nano Banana for creative)
07:10toolGemini 3 Flash (auto-assign classifier)
08:40productClaude Code Magic Course
09:20toolGoogle Workspace CLI (for Gmail drafting)
10:24toolcron (Unix scheduler)
18:00toolSupabase / Neon / local SQLite (memory storage options)
20:20toolObsidian graph view (visual inspiration)
20:40toolLoom (screen-record your wishlist into Gemini)
20:40toolGemini video understanding API
Lines you could clip.
00:16
“This is one small element that's called the hive mind, which is essentially a shared memory state of my ever growing team of agents.”
02:58
“Whether you want to use Claude code, codex, or even both, you'll be able to apply all the principles I'm about to show you to whatever LLM you want.”
03:24
“Your computer is the engine. Telegram is the steering wheel.”
24:14
“If your back of house is organized, then everything else is purely a cherry on top.”
26:40
“The deeper you get into this agentic OS or AIOS or whatever the term will be five hours from now, the deeper you understand that this is a data engineering problem. This is not an AI problem.”
27:20
“There's no Hermes agent or OpenClaw agent configuration that's gonna be perfect for you. This is an iterative process.”
00:56
“With just a few prompts and slash commands, you can have your own LLM council at your fingertips.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphorstory
00:00So we're looking at right now is something that I practically use every single day to run my business. And this is one small element that's called the hive mind, which is essentially a shared memory state of my ever growing team of agents. Agents.
00:13Each set of agents correspond to different parts of my digital brain and each one of these little nodes here represent tasks that they've completed. So if I want an overview of the overall activity of all my agents and all the knowledge that they know about each other, this is one of the many ways I can represent that. Now if I want something more on the two d side, something very similar to Obsidian's graph view, then we have this portion where we can filter on specific agents, and then we can even search specific tasks.
00:41So if I search something related to Gmail, you'll see every single node that remains that represents a task where Gmail was executed. And if I wanted a list view, I could go through this entire table and see each and every action of every single agent in the same place. Let's say I was feeling particularly introverted today and I wanted to have the equivalent of an .
01:04I can go to this specific section called my war room and instead of using my voice, I could put on my headphones and go straight to the chat room. And when I click on text meeting right here and we click through, this will open up a brand new chat room where I can use all of my specific commands to have stand ups and discussions with all of my agents.
01:23So all I have to do is I can go into maybe small mode or large mode if I wanna be able to see every single agent at my left hand sidebar as well as a status that shows their last completed task. And I can go to the very bottom here, and I can do slash stand up.
01:38And I can either tag specific agents to give me the status report or I can have them all tell me what they've done in tandem. So if I just send this specific command right here, each one of my agents will be invoked.
01:51It will go through its associated database and set of memories to tell me exactly what has completed in the last twenty four hours. And you can see each one of them goes through and gives me a very succinct status report ending with our meta agent.
02:04So we have the research agent chime in and tell me that this part of the YouTube space might be a little bit saturated. Then we have comms who is my designated content agent, the one I probably wanna listen to the most. But each one will chime in.
02:17The main agent will kind of look at the other responses and give me the cohesive response on behalf of the rest. So with just a few prompts and slash commands, you can have your own .
02:29Now big picture, I wanna be very clear about what this video is. It's not meant to be a step by step walkthrough on exactly how you can go from absolutely zero to this. And the reason is this takes hundreds of hours to start, iterate, and refine.
02:43It's meant to show you what's possible, how you can get there, and most importantly, how to break down this very opaque concept of this AI operating system. Whatever the term of the day is, this is purely a data organization exercise with some layers on top.
02:58So in the next twenty ish minutes, my goal is to show you what I use this for day in and day out, how you can set it up for yourself, and most importantly, breaking down this very opaque concept of the AI operating system into the simplest terms so you can finally understand it and implement it on your own. Whether you want to use Claude code, codex, or even both, you'll be able to apply all the principles I'm about to show you to whatever LM you want.
03:22Let's get into it. And before we deep dive into every bell and whistle that this system has to offer, it's important to level set for anyone that's new here or hasn't watched my prior videos on this. A few months ago, I was quickly jaded with what OpenClaw could and couldn't do, and all I wanted was an easy way to connect my existing Cloud Code subscription to my existing Cloud Code ecosystem and connect them together.
03:45And I went from wanting to creating a way where you could use a bridge called the Anthropic SDK and connecting it to your ClaudeCode system so any skill, any project, any plug in was accessible through my Telegram interface.
03:58And I went from interacting with one singular agent to a team of agents to creating a war room where I had an experimental feature to go back and forth using my voice. And now I've evolved this hive mind to allow me to interact with all of these specialized knowledge, the subject matter expertise of every single employee in my workforce.
04:18And the best part of this system is because my Cloud Code is already connected to all of these services through various CLIs, integrations, and skills, all of my agents inherit this existing infrastructure.
04:30Now let's make this practical by walking you through one actual business case. So let's say I have a Meta agent, and this is something that I've been experimenting with recently. I'm a novice when it comes to Meta ads, so I wanted the power of Cloud Code to connect to services like the Meta command line interface, which is a brand new way to look at, track, and understand ad performance.
04:51Behind the scenes, it has a set of instructions walking you through exactly what its role is and how to interact with all of these services, And underneath the hood, this is how it works. When I ask it a question, it will go and proactively search the Meta Ads command line interface. This is a fancy way for Claude Code to go and investigate the entire API and see every single part of a campaign, ad spend, ROAS, and a bunch of other words that up until a week or two ago, I had no idea how to actually apply.
05:21And then because I want the information distilled and portrayed in a very specific way to be displayed on my mobile device, I create a meta ad skill and this basically creates a custom report. Once the skill exists globally, every single one of my agents automatically inherits it.
05:37And because of this, I can now create a scheduled task that will send me a ping every morning at 07:30AM of a full summary of ad performance. And if I wanted to go back and forth with my agent and get to the bottom of which ad is working, are any working?
05:52Do I have to create some brand new creative? I could also task it with creating that creative using the Gemini skill to use something like Nano Banana to create the creative, add the creative, add some money, put some spend on it, and see how it performs. And this is one small microcosm of every single task that I give my set of agents so I can focus on the things that matter.
06:13And this is what it actually looks like. So today, I got this report. It breaks down the spend, the actions, a couple blind spots that I might have, and each one of these hyperlinks right here linked to the direct ad that it's referring to.
06:26And again because this is more than just a reporting tool, it can give me a quick take at the bottom here, tell me if I have a winner, a loser, what I'm doing wrong, and how x is affecting y in a way that I might not know out of the box. Now in a similar fashion, I'm gonna go through the relevant tabs that you might wanna adapt to your own system and show you one, how it works tactically, and underneath the hood, what's involved to make that happen.
06:49So this mission control dashboard is meant to be a glorified Kanban where you have all of your agents in one place. You can spy and see if any tasks are happening in parallel real time, and you can switch the layout as you wish, and you can always create a new task and drag and drop it to the agent of your choice.
07:07And if you don't want the cognitive load of deciding which agent should take this specific task, you can always do an auto assign. So in this case, if we take this task, which is purely sending Mark a message saying hi, you can click on auto assign, and then this will use Gemini underneath the hood using the cheapest model from Gemini, which is inconsequential from a cost perspective to tell you which agent is best geared for this.
07:32And naturally, it chooses the comms agent. So this will now be queued and then sent over and within seconds to a minute depending on the size of the task and the length of the task, it will be executed and you can follow-up on Telegram. And by the way, if you wanted to swap Telegram for Slack or Discord, you can do that.
07:50You just have to set up the connection and make sure it's stable. And within seconds, the system gets pinged that there's a queued task, it starts running, and then it should execute until it's completed. And by the way, as usual, I'm gonna give you a series of resources to help you reverse engineer as much as I'm showing here as possible.
08:06But if you want my carbon copy system that I keep on updating every single week with brand new features, adjustments, and everything needed to make this awesome, then you'll wanna check out the first link in the description below. Not only do you get access to this current version of Clawd Claw, but we're working on multiple versions.
08:22One for enterprise and eventually for local, and on top of that, you get access to myself, my team of coaches, as well as the Claude code magic course, which is our living course for anything related to Claude code. If that sounds interesting to you, then check out the first link in the description below, and let's get back to video.
08:39And because my comms agent is a subject matter expert at anything related to communication including my email, my Outlook, my LinkedIn, etcetera, it knows that for Gmail to use the Google workspace command line interface, which is already a part of my Cloud Code system and leverage it to draft this email. Now if we take a peek behind the curtain, all that's happening is that we have a series of tasks here and I'm using the Gemini skill using the absolute cheapest model Gemini three flash to take all the tasks and auto assign it with a very basic system prompt.
09:10And the system prompt is dynamic. Based on all the agents that we have, go and decide what is the best geared agent and here's a description of all of the agents. So then Gemini classifies it, it adds it to the queue, and then the system gets pinged like you saw.
09:24It starts running, and once it's done, we have completion. And the hardest part of the mission control is making sure to remember that the front end should always have a perfect symbiosis with the back end so that when you have a task on screen that propagates to something like your telegram API, your different APIs are being used, and making sure that this whole system is cohesive.
09:44Now the schedule tab is very straightforward. It purely creates what are called cron jobs and these are scheduled jobs that run on your computer locally. And even if you host it on a VPS or a cloud provider, you can schedule it from there as well.
09:57And if we click on something like this schedule task for our meta ad CLI, you could see we could just add a time from here really easily. We could say every weekday, every weekend, custom, and anything like an nnn or a make.com experience, this is the exact same thing. The trick with this one is if you want the front end to be very straightforward, this is what cron looks like underneath the hood.
10:17It's a series of numbers and characters that denote things like cadence, frequency, day, year, etcetera. You don't want to see that as a front end user.
10:26So the only thing you would want to tell Claude code is instead of showing me this raw input, show me the equivalent raw input as English or French or whatever language that it is that you speak. Now the agents tab acts as the command center for all of your existing and brand new agents. For your existing ones, you can switch the model right here Assuming you're using Claude Code and you're using your existing Claude Code subscription, you can switch the model and this will propagate, fancy word for automatically update anywhere else that this is used.
10:55If you wanna change the personality, if you wanna change the task list, if you want to tell it to focus on this skill versus that, one, you could just tell it through telegram or you could just write it directly through here. And you can always stop your agents, delete them, restart them, and I even added a feature that you might want to steal from me which is having suggestions on agents that deserve to exist that don't.
11:19And a good way of doing this is you would have some language model. In my case, I don't want to sacrifice my precious Claude code tokens so I will put my Gemini three flash API on the front lines to look through all of my conversations, scan through and see what agents am I overburdening.
11:36And in this case, you can see that comms is doing way more than it should. It's handling WhatsApp. It's handling school.
11:43It's handling mail. It's handling every single thing you can imagine. So maybe it makes sense to have another agent that's called email manager.
11:50And the best part is if you want to create a new agent, you should design it. So when you click on new agent, it lets you just write the name, the display name, the description, and it gives you the link to the bot token.
12:01So all you have to do is go to Telegram, create a brand new bot, you take the existing token, you copy and paste it into your front end, and then you should be able to activate and ping it from here. And underneath the hood, you can add way more sophistication than I do, but every single agent is purely a CloudMD file and what's called a YAML file.
12:19This is a breakdown of its configuration. You could add its own specific rules and skills completely specific to this agent, or you can keep it nice and minimalistic and just layer on top of your existing global skills in your system.
12:32And when it comes to the suggestions feature, like I said, all it's doing is looking at your usage over time, looking for really used and abused agents, looking through your conversations, and looking for patterns, using pure natural language.
12:46Because when you use things like the Gemini models, you have enormous context windows, very cheap inference, so you can afford to take entire JSON files of your conversations and automatically have that sift through and decide what makes the most sense.
13:00When it comes to the chat tab, this is just designed to have all of your communication centralized in one place. So if you wanna keep tabs literally on all of your agents and see all of your conversation history with them and continue the conversation at the very bottom of here, like you can write status update, what's next, you can type whatever message you want, you could just say hi, and behind the scenes, it will still use your existing harness, your existing subscription just from this online portal.
13:26And assuming you use this with multiple platforms, this can help distill all of those conversations into one unified view. Now when it comes to memory systems, everyone has an opinion as to what is the best.
13:38And I even dropped a video a few weeks ago walking through the different levels of memory so you can decide for yourself what makes the most sense for yourself, your business, and your day to day. In my memory world, I have five to six different layers and these layers really roll up to these three categories which are importance, salience, and recency.
13:57And the one thing that I can do in this blurred memories tab is I can write something like let's say Gmail and this will show any memory related to that. So you can make your , which when it comes to memory, it's beyond just throwing a vector database and throwing everything in there and just expecting it to work because memory is not just about the setup, but it's also about maintenance.
14:19And one additional thing that you can do is you can create insights on your memory. So you can have, again, a language model, a cheap workhorse, take a look and scan through your memories and derive insights about you, Things that you don't know about yourself.
14:32And this would be the equivalent of doing slash insights in Claude code to get the full breakdown of all the pros and cons of how you've been using it over the last thirty days. Now, I've covered the entire deep dive of how I've structured memory in a prior version of this video, but the TLDR is this. You want Claude Coe to interview you on how you wanna deal with fresh memories and how you wanna deal with fading memories and important memories.
14:55Do Do you want important memories to be pinned forever? Do you want memories that fade to fade into non existence, or do you want them to be stored somewhere? And when it comes to the semantics of this system, pun intended, you have different factors that you'd wanna consider.
15:09One of them like I said is salience, is a fancy word for the importance of a memory and whether you want to create your own embeddings. Now the beauty of a system like this is that you can have it run on an online database, something like a Supabase, a Neon, or like myself, all of this is contained in a local SQLite database.
15:28So each and every part from who the agents are to all the conversations, to all the scheduled tasks, to all of my memories, to everything in the hive mind is all stored for free on my local database. So the sky's the limit, and you can make this work irrespective of the platform of your choice. Then to help you build and design your own memory system, I'm gonna give you my memory skill on top of a series of other things that you can use to interview yourself and have Claude Code build that system for you.
15:56Now when it comes to the holy grail of the hive mind, this three d version and the two d version really are based on the foundation of the list view. If the list view is actually working and the list view is operational and all the agents are logging in real time and all of the database is set up, everything else is just additive.
16:14So once you set up the boring stuff and you make sure that this table is always populated and it's populated with the latest and greatest summaries of what that agent is doing, you can always layer on the two d version or the three d version, whatever makes the most sense for you. And when it comes to implementing these more fancy features that are easy on the eye, but tricky on the back end, this is my mental model of how I create them.
16:36So in this case, I was inspired to create for myself the very thing that I really like about Obsidian, which is this graph view where if you're not familiar, each one of these bullets denotes a project, a task, or something similar. So all I did was actually screen record using Loom. You could use things like Tele, whatever screen recording of your choice, and I went through and I told it.
16:58I basically spoke freestyle. This is what I love about the Obsidian Graph View.
17:03I just wish that I could create it in a way that really synchronizes in real time to what my agents are doing. And then all I did was I hooked up to my Claude code system, this skill, which is the Gemini skill, and as a part of this skill, if you zoom in here, one of them is processing video through their video understanding API.
17:23So all I did is take the Loom, download it, drag and drop the file path and I said, hey, want to build this, but I wanted to integrate into my mission control in this way. And one added thing is depending on your system, this three d version again looks beautiful, but it uses a lot of resources. So if you don't have the best of computers, then the two d version is the most efficient.
17:44Now ending off with the beloved War Room, in my last video, I walked through at length on how I designed this voice and live meetings capability. The voice being the part where I can launch this specific room, it plays a beautiful tune behind the scenes, and it allows me to interact with all my agents synchronously. And I went from just having the ability to speak to my agents in this way to now having the ability to go back to the war room, go to the picker, and then I can select who should always be in our stand ups when we do our daily meetings.
18:13And the way I designed this theoretically is very simple. I wanted to not only always have a synchronized list of my latest and greatest agents, and I also wanted to have a list of their latest and greatest tasks according to the hive mind, but I also wanted to create a special set of slash commands specific to this room.
18:32And a couple of them you already saw, one of them is slash stand up, and one of them is slash discuss, but it can also pin an agent so that they lead every single reply. So if you wanna have your council meeting, but you wanna designate one specific agent as the core leader of this meeting, then you can do so. And then everything else is just clearing the agent history, and the hardest part of this is making sure that when you have this conversation, each agent is aware of the last response and the overall context.
19:00And luckily, I'm not a rocket scientist. I just asked Claude Code, can you find a way that in this specific ecosystem of the War Room that we can have this conversation and it all has context on all the other replies as well as their underlying Claude MDs, YAML files, etcetera.
19:17And on top of that, if you played around with things like custom GPTs back in the day, then you'd remember you could tag specific GPTs in the exact same manner. If I want to talk to my content agent and ask it, give me advice for creating a good video on a Claude Claw system for YouTube, this will automatically tag that specific agent.
19:40And again, these are small things, but everything is iterative and everything is additive. So the core question to ask yourself is what is the eighty twenty for you in terms of features that actually drive business value and drive ROI versus something that's purely a novelty? And behind the scenes, the stand up and discuss slash commands are purely pinging this non visible prompt to you and I that's basically telling them, give me a status report on what you've done according to your entries in the hive mind as of the last twenty four hours.
20:08So this is really where you see the entire system coming together and acting cohesively. Alright. So now that you have an understanding of how my system works, I wanna walk through this paradigm of the AIOS so you can fully understand it and see how it plays into what I've shown you thus far.
20:23Now the main thing I'd say about this system is that all of this top layer here, the part where you have the fancy dashboard that I showed you, all the functionality and features, all of that is cute. But the part that really matters, the foundational layer is this bottom layer.
20:38How is your data organized? If your back of house is organized, then everything else is purely a . So if your system looks like this when you open up your desktop and you have files and folders all over the place of different file types and sizes and you expect that even taking my carbon copy system and layering it on top of this will give you something very powerful, you'd be wrong.
21:01The art of making this work is really the art of just being hygienic with your files, your skills, and deciding very proactively what deserves to be a global skill and what deserves to be a project level skill.
21:13And again, skills are universal whether it's Codex, Claude, or Gemini. And beyond that, it really just comes down to the different skills, rules, and command line interfaces and integrations they make available to existing OS.
21:26And when it comes to the agents, the skills in the CloudMDs, you again just have to decide what deserves to graduate or be promoted to a global skill that every single agent should have access access to versus what should I pigeonhole, make specific to one particular terminal agent or project.
21:44And once you have this bottom layer figured out, everything else is easier. You just add on some form of memory. You would decide what kind of database makes the most sense for me.
21:52Do I care about scheduling? Do I care about specific agents? Do I want specific skills for these agents?
21:58Or do I purely wanna just build on the existing skills in my system? Once you make those decisions, the last part is how would you like to interact with your system on the go? In my case, it's Telegram because the blast radius of my business and my business contacts and my personal life is zero.
22:15In your case, it could be signal. It could be discord. It could be whatever.
22:19So the deeper you get into this agentic OS or AIOS or whatever the term will be five hours from now, the deeper you understand that this is a data engineering problem. This is not a AI problem. And once you get the hang of one of them and you start playing around and giving it feedback saying, you know what, I hated the way you did this.
22:38This is where you don't have self improving, but you have iterative improving, which when it comes down to it, for business purposes, the best way to build agents is to have an where there is no one solution. There's no Hermes agent or OpenClaw agent configuration that's gonna be perfect for you.
22:56This is an iterative process, and basically fixing your and layering this on top allows you to go iteratively and decide is this a data layer problem or is this my operating system problem? And to help you not only understand, but implement a lot of the concepts that I walked you through, I'm gonna make available to you a series of resources.
23:14This includes a full blueprint of my existing system that you could feed to Claude code, codex, and reverse engineer everything that I showed you here. You could also take the transcript of the video. You could even take the video itself and feed that to Gemini and ask it to create the perfect prompt for you to design it, but also give you as many supporting documents as you need along with the skills to create the memory system of your dreams.
23:36And again, this is something that I've been working on nonstop. I've been iterating on it for months, and everyone in my exclusive community has had access to each and every version and all the supplementary So if you want the carbon copy version of my command center and you wanna keep tabs on everything else we have coming up, which is even an enterprise grade version of this system, then you'll wanna check out the first link in the description below, and I'll see you in my early adopters community.
24:01And for the rest of you, if you found this video helpful, inspirational, and hopefully informative, then please let me know by leaving a like on the video and a comment. It really helps the video and the reach, and I'll see you all in the next one.
The Hook
The bait, then the rug-pull.
Mark opens not with a promise but with a flex — a glowing 3D brain of agent nodes rotating in real time — and tells you straight up this is what he uses every single day. The hook is the artifact: a hive-mind visualization you've never seen anyone else ship, with a Telegram-powered war room sitting one tab away.
Named ideas worth stealing.
03:24model
The Wrapper / Bridge / Engine
- Telegram = the steering wheel (primary door)
- Anthropic SDK = the bridge
- Claude Code on your local computer = the engine
- Hive Mind = the wrapper
Mental model for the whole architecture. The remote control (chat app) is intentionally separate from the brain (Claude Code) so you can swap front-ends without rebuilding the back-end.
Steal forany local-first AI tool you want to control from your phone — Joe could literally apply this to MCN agents or JoeFlow.
16:50concept
Three memory categories (rolled up from 5–6 layers)
- Importance
- Salience
- Recency
Compresses every memory-system decision into three axes — what matters, what's resonant right now, and what's recent. Pinned vs. fading vs. archived are the user-facing knobs.
Steal forany agent memory store — replaces ad-hoc vector-dump approaches.
24:14model
AI OS stack — foundation up
- 1. File / skill / CLI hygiene (back of house)
- 2. Agents + CLAUDE.md + YAML
- 3. Memory layer
- 4. Scheduler
- 5. Remote-control surface (Telegram / Discord / Signal)
Build bottom-up. The fancy dashboard is layer 5; if the bottom is messy nothing on top will save you. Most people fail because they layer features on disorganized files.
Steal forany creator pitching an 'agentic OS' product — the order matters and most demos lie about it.
07:13concept
Auto-assign pattern (cheap classifier + expensive worker)
Use Gemini Flash for the routing decision (free-ish), reserve Claude Code tokens for the actual work. Same pattern shows up in the suggestions feature and the insights feature.
Steal forany agent system where routing is half the cost of doing — JACE/REESE/SAGE/RYDER orchestration.
26:40concept
Infinite-game mindset for agent design
There is no perfect agent config. Build, ship, give feedback ('I hated the way you did that'), iterate. Not self-improving — iteratively improving.
Steal forJoe's stance on Mod Producer / JACE / etc. — frame agent products as living systems, not finished SKUs.
How they asked for the click.
VERBAL ASK
08:24product
“If you want my carbon copy system that I keep on updating every single week with brand new features, adjustments, and everything needed to make this awesome, then you'll wanna check out the first link in the description below.”
Two-pass CTA: a mid-roll sponsor break around 8:00 pitching the skool community + Claude Code Magic Course, then a softer end-of-video repeat at 23:00 framed as 'resources to reverse engineer everything.' First link is intentionally weighted — the entire description's chapter list works as a teaser that pushes viewers to the artifact.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
