32:39NetworkChuck · Tutorial
I'm switching to Hermes (goodbye OpenClaw!!)
A 32-minute live walkthrough where NetworkChuck installs a self-improving AI agent, names it Ron Weasley, and never looks back at OpenClaw.
May 20thThe argument in one line.
A meta-harness that turns every AI agent, regardless of which company built it, into an employee inside one org chart with tasks, budgets, and human approval gates solves the coordination problem that running five separate chat sessions never could.
Read if. Skip if.
READ IF YOU ARE…
- You're already running more than one AI coding or automation agent and have no shared way for them to hand off work.
- You manage infrastructure (network, storage, security) and want an always-on team of agents investigating incidents instead of one-off chat sessions.
- You want a concrete example of human-in-the-loop approval gates on autonomous agents before you build or adopt one yourself.
- You're curious how a 'meta-harness' differs from a single agent framework and whether it's worth adding on top of tools you already use.
SKIP IF…
- You want a code-level tutorial on building an agent framework from scratch. This is a product walkthrough, not a build-it-yourself guide.
- You only use one AI agent for occasional one-off tasks and have no coordination problem to solve yet.
The full version, fast.
Paperclip is an open-source meta-harness: instead of replacing Claude Code, Codex, or Hermes, it sits above them and turns each one into an employee inside a single company, with a CEO, an org chart, assigned tasks, budgets, and human approval gates. The host installs it on a VM, hires a Claude Code CEO named Dumbledore plus three existing Hermes agents, a Codex security agent, and two local-model workers, then assigns the whole team a real bug: his studio NAS drops every time someone flushes the toilet. The agents investigate in parallel, produce a shared report, ask him decision questions along the way, and land on a root cause that has nothing to do with the toilet. He also sets up a scheduled 'routine' for recurring checks and connects a dark-web monitoring tool through an agent-held API secret.
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 – 01:06
01 · The problem with having lots of agents
The host names the tools he already juggles (Hermes, Codex, Claude Code) and introduces Paperclip, a meta-harness, and its creator Dotta as a way to make agents from different vendors work together as one company.
01:06 – 02:10
02 · The battle plan and where to run it
States the real test case: the studio NAS drops every time someone flushes the toilet. Covers running Paperclip on a local machine (Mac/Linux/WSL2) versus a dedicated VM or VPS for anyone who wants to commit to it long-term.
02:10 – 03:39
03 · Hiring the first employees
Introduces three existing Hermes agents, Ron (CTO), Fred (network engineer), George (storage engineer), sketches the org chart, and explains that a Claude Code CEO will sit above them. Frames the whole exercise as 'harness maxing': any new model or harness can be brought in without rebuilding the company.
03:39 – 05:56
04 · Sponsor: Flare and the dark web exposure check
Flare, a dark-web and identity threat intelligence platform, is shown scanning the host's own exposure: leaked credentials, chat messages on criminal forums, and infected devices tied to his academy's members. Includes the stat that most breaches start with stolen logins, not a break-in.
05:56 – 08:48
05 · Installing Paperclip and naming the CEO
Installs Paperclip on an Ubuntu VM with a two-command script, names the CEO agent Dumbledore, chooses Claude Code as its underlying harness, hits a setup error from Claude Code not being installed yet, fixes it, and walks through the onboarding flow for the first time.
08:48 – 11:59
06 · Building the org chart
Explains the board/CEO/employee hierarchy, approves Ron's hire, tests agent-to-agent communication by having Dumbledore ping Ron directly, then hires Fred and George and fills out the rest of the team (Arthur Weasley on Codex, Watchdog and Help Desk on Pi, Mad-Eye Moody as security reviewer).
11:59 – 15:51
07 · Assigning the toilet ticket
Creates a project for the network issue and hands the full ticket to Dumbledore, who opens a case, signs out sub-tasks to every agent in parallel, and starts producing shared artifacts like a studio network map while the host watches progress from the dashboard.
15:51 – 18:42
08 · Decision gates and the shrinking suspect list
The agents rule out storage and isolate a specific switch signature, then surface human decisions: whether to run a controlled flush test, replace a suspect fiber optic, or audit a different access path. The host picks options through the approval interface rather than a plain chat reply.
18:42 – 21:37
09 · The verdict: it isn't the toilet
Dumbledore delivers a written verdict ruling out the NAS, router, and spanning tree, then identifies four fiber links on one switch that have been silently flapping for fifteen months, with one port logging over 20,000 link-down events against 123 on a comparable port.
21:37 – 23:41
10 · Root cause and the physical fix
Traces the failures to a batch of cheap third-party SFP transceivers bought off Amazon with a roughly 50% in-service failure rate, then swaps them for official MicroTik modules on camera.
23:41 – 25:34
11 · Routines and connecting Flare
Sets up a recurring 'routine' that checks Ceph storage health every morning at 10am and flags low capacity as a decision, then hires a security agent named Filch on Claude Code Opus 5, gives it Flare API credentials as a scoped secret, and runs a live exposure sweep.
25:34 – 27:30
12 · Standups, exports, and the close
Shows the daily standup routine where agents report on their work and question each other, the org export/import feature for taking a whole company elsewhere, and closes with a call to try Paperclip and comment with other harnesses to test.
Lines worth screenshotting.
- A meta-harness doesn't replace Claude Code, Codex, or Hermes, it sits above them so agents from different companies can work inside the same org chart.
- Giving agents a company structure, a CEO, managers, and direct reports, produces more useful coordination than an open 'agent town' where bots just talk to each other.
- Agents in this system escalate blockers to their manager first and only surface to the human board member if the manager can't resolve it.
- A pre-trained remote agent that already lives on its own VM with domain skills can be 'hired' into a new company instead of built from scratch.
- 54% of ransomware victims had their credentials sitting in info-stealer logs before the breach, and most companies take 36 hours to notice while attackers need only 48 minutes.
- Session cookies stolen by info-stealer malware let an attacker skip both the password and two-factor authentication entirely.
- One compromised device can expose credentials for hundreds of thousands of other accounts because info-stealer logs get resold and aggregated on criminal markets.
- A 'routine' is a scheduled task an agent runs on its own, like a daily storage health check, and it can escalate a decision to a human the same way an on-demand task can.
- Human approval gates in the system aren't a rubber stamp; the host is shown multiple concrete options (run a controlled flush test, swap the part now, audit a different path) and has to pick one.
- An 'artifact' is any file or report an agent produces, stored in the company and traceable back to the task and the agent that made it.
- The agents ruled out the obvious suspects, the NAS, the router, and spanning tree, by producing a written report with per-port failure counts before proposing a fix.
- The actual root cause was a batch of cheap third-party SFP transceivers with a roughly 50% in-service failure rate, not the toilet the whole video is framed around.
- One port in the faulty batch logged over 20,000 link-down events against 123 for a comparable port on the same switch, model, and firmware, which is what made the bad batch provable rather than a guess.
- Exporting an entire agent company, its agents, skills, routines, and tasks, into a portable package means the setup isn't locked to one machine or one run.
- Connecting a new API credential to a single named agent (rather than the whole company) limits which agent can act on that integration.
A shared org chart beats a pile of separate agent chats.
WHAT TO LEARN
Coordinating multiple AI agents gets easier once they share a company structure, a reporting chain, and human approval gates, rather than each living in its own disconnected chat window.
- A coordination layer that sits above individual agent tools, rather than replacing them, lets you keep using whatever harness or model already works for a given job.
- Giving agents a manager to report to, instead of leaving them to talk freely to each other, keeps escalation focused and stops idle cross-talk.
- A pre-built agent that already has domain skills and lives on its own machine can be added to a team instantly, the same way you'd bring on a specialist contractor.
- Most account breaches start with a stolen password or session cookie, not a technical break-in, and most companies take far longer to notice than attackers need to act.
- Session cookies stolen by info-stealer malware can skip two-factor authentication entirely, so credential rotation alone doesn't close the exposure.
- Presenting a human with a short list of concrete options, rather than an open question, makes an approval checkpoint faster and less likely to be rubber-stamped.
- A written report with specific per-unit failure counts is what turns a suspected cause into a provable one, versus a plausible-sounding guess.
- Recurring automated checks (a daily health scan) catch slow-building problems, like shrinking storage capacity, before they become urgent.
- Exporting an entire working setup, agents, tasks, and schedules, into a portable package is what keeps it from being locked to one machine or one session.
- Scoping a new credential to a single agent, instead of the whole system, limits the blast radius if that one integration gets misused or compromised.
Terms worth knowing.
- Meta-harness
- A coordination layer that sits above individual AI agent tools (Claude Code, Codex, Hermes, local models) so agents from different vendors can share one org chart, tasks, and budgets instead of running as separate, disconnected chat sessions.
- Hermes agent
- A pre-configured, always-on AI agent that lives on its own virtual machine with a fixed set of skills, brought into a new company ready to work rather than trained from a blank prompt.
- Case
- A grouped set of related tasks an agent opens to investigate a single problem, letting multiple sub-agents work different angles of the same issue under one umbrella.
- Routine
- A task an agent runs automatically on a recurring schedule, such as a daily storage health check, instead of being triggered manually each time.
- Watchdog agent
- An agent role dedicated to reviewing whether a task that another agent marked finished or stopped is genuinely complete or actually blocked.
- SFP transceiver
- A small pluggable module that plugs into a network switch port to convert electrical signals to fiber-optic light, letting the switch connect over fiber cable.
- Link flap
- A network fault where a port's connection repeatedly drops and reconnects in quick succession, rather than failing once and staying down.
- Info-stealer malware
- Malware that harvests saved passwords, browser session cookies, and multi-factor tokens from an infected device and sends them to an attacker, often without the victim noticing.
Things they pointed at.
Lines you could clip.
01:45
“No matter what new harness comes out or model, Paperclip is a meta harness. We can bring any other harness or model to it.”
05:05
“Attackers don't hack in. They log in.”
19:40
“It's not the toilet, and it is not a mystery.”
21:08
“Port 5 has logged 20,052 link downs since the switch booted. A comparable port has only done 123.”
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.
So we're all using agents, right? Lots of agents. Hermes, Codex, Claude Code, OpenClaw.
In my company, I have eight. It's great, but managing them, that's hard. How does everyone talk to them?
Can they talk to each other? Who does what? Who's keeping track of that?
And what about security? We've seen attempts at this, but it's really not it. But I think I found something that is.
It's called Paperclip. It's a meta harness. You bring Claude Code, Codex, Hermes, OpenClaw into Paperclip.
By the way, this is Dota, creator of Paperclip. This is also the first time he has shown his face in public. Our first interview looked like this, but he's pulling a dream so he can put a personality behind this project.
In Paperclip, your agents become employees. You give them managers, you assign them tasks. They can talk to each other, but it's not what you think.
It's through the task, intentional and focused. I haven't seen a lot of super value out of like... an agent conference room or an agent town where you just set your agents to sit there and talk to each other.
And it doesn't matter what harness you have. Everyone can be brought into this company. Isn't that amazing?
And by the way, my agents have a standup with each other every single day and it's awesome. So in this video, we're actually going to solve a weird problem I have. So here in Network Chuck Studios, anytime someone flushes the toilet, everyone is disconnected from the NAS.
Me, the editors, I don't know. It's so strange. So to solve this, we're going to do something I've been trying to do for a long time.
Set up my agentic IT department. All AI agents, network and systems engineers, storage, security, a legit company running on paperclip. Get your copy ready.
Let's go.
Okay, we're setting up an AI company from scratch. Here's our battle plan. now listen even if you don't have a company you want to start this is actually pretty fun to play with with paperclip you do need something to run it on where are you going to put it In most cases, if you just wanna try this out, the computer you're using right now, if you're using Mac or Linux or WSL two on windows, that's probably the fastest way to get spun up and just start playing with it.
If you wanna commit, do what I'm doing. Put that sucker on its own server. This could be a virtual machine like on Proxmox or a VPS in the cloud.
I'm gonna show you both. And by the end of this video, you're gonna have your very own paperclip company running agents. You want that, right?
Yes, you do. Let's take a sip of coffee real quick. All of us together.
You ready?
Now here's my absolute favorite part about this for my employees, from my it department, I'm bringing in three Hermes agents, Ron, the CTO, Fred, the network engineer and George, the storage engineer for obvious reasons. Now here's why this is amazing. These Hermes agents they're pre -trained.
They already exist. They live on their own VMs and they already have every skill they need to help me solve my problems. They are fully functional employees.
I just hired them off LinkedIn essentially. So we're gonna set up paperclip. We're gonna hire our employees.
which just sounds so fun, which the, uh, org chart will be like this Ron to CTO, Fred and George report to him. And then just above Ron, I'm actually gonna hire a CEO, which I'll actually have running cloud code and not Hermes. And just for fun, I'll hire some more employees too.
Probably a security engineer running codex. And let's say a help desk engineer running a local model. Once we've hired our team, we will then assign a task, a task to solve our toilet problems.
And we'll watch them go to work. Now I want you to pause here for a moment and think about what we're doing here. We're harness maxing.
No matter what new harness comes out or model paperclip is a meta harness. We can bring any other harness or model to it. Paperclip can use it.
Paperclip is our company. And it's a company that can run the greatest and latest models and the greatest and latest harnesses and all the features they come out with. Just think about that.
Doesn't that make you excited to try this? It does coffee break. Okay.
So this it department's gonna take care of what's inside my network. But what about outside? What about things that are already exposed about me and my company?
Have I already been hacked? That's where the sponsor of this video flare comes in. Don't you dare click away.
This is actually really cool. I know you're tempted. Get your finger off that right now.
Take a sip of coffee and enjoy some cool stuff. Flare is an identity focused threat, intelligent platform. It's gonna watch the dark web.
It's gonna watch telegram, the hacker forums. And it tells you if your company's stuff has been out there, credentials, access sensitive files. And, uh, I plugged my stuff in.
They let me try it. Oh, It's so bad.
Like I'll be the first to admit. I probably don't have the best op sec, but I'm not that bad, right? No, I am.
It's bad. Can I, I'm gonna show you a little bit of it right here. Check it out.
76 leaked credentials, ransom leak, 390 chat messages, 23 lookalike domains, 32 ,000 infected devices. Now what does that mean? I didn't know at first.
It's actually crazier than you think. This infected devices thing is actually real people. People who have signed up for my academy.
and their credentials have been leaked. Now it's not like my academy got hacked. The academy is fine, but it's that these people and their credentials have been stolen.
They're sitting on Russian markets and telegram right now, 10 bucks a log. One of those machines have 300 ,000 safe passwords in it. It found telegram bots with cookie sessions from my academy people posted on the dark web, which means you don't need credentials.
You just use those cookie sessions, skips two factor. So listen up. We know how breaches happen now.
It's not a mystery. Attackers don't hack in. They log in.
Info Stealer malware grabs everything. Passwords, session cookies, MFA tokens. 54 % of ransomware victims had their info sitting inside Stealer logs way before they ever got hacked.
And most companies take 36 hours before they even noticed something has happened. Attackers only need 48 minutes. But with Flare, You realize things happen within seconds.
They have real time detection and actually checks against intra ID, whether an account is actually real and active, and it can kill a session before anyone opens a ticket. Look, I'm a business owner. I know how it feels to be sitting here wondering, am I about to get hacked?
Have I already been hacked? What's going on out there? Try flare out to see where you're at, get a baseline, see how they can help you.
You can try flare for free at my link below, and you'll get a chance to see what attackers see. So massive shout out to flare for sponsoring this video. And also what I'm excited to try here in a moment, I'm gonna connect this sucker to paperclip.
There's an API. I'm gonna have my agents handle this stuff for me.
Step one, let's install paperclip and it's gonna be fast here. I have my Ubuntu VM and I'm gonna install it with two commands and it should be pretty similar. Whether you're on Mac Linux or WSL two detailed steps below first, we'll download the installation script and then second we'll run it little coffee break while it's running.
Now because I'm installing Paperclip on a VM, I'm doing the advanced setup. For most of you, Quickstart is the way to go. So I'm gonna jump through a few of these settings.
I'll make sure I can access Paperclip over my local network without any issues, and then get that bad boy started up. Okay, our install is done. Now for you, you'll either go out to local host or your IP address, or maybe your domain name if you set it like me, and make sure you're going to port 3100.
And then you'll onboard your company and set up your first account.
What should we name our CEO? Dumbledore, duh. But for my CEO, I want to run it with Cloud Code.
Now keep in mind paperclip itself is not an agent harness. It's a meta harness. So as we set up cloud code or codex, that's gonna actually be running on that paperclip server.
Or if you're installing paperclip on your machine right now that you're using and watching me on, it'll run on that machine. So I'll choose cloud code per our plan. It said ready, but I got a bunch of error messages.
What happened? I don't know. We'll figure it out in a second.
But look how cute Dumbledore looks. Oh my gosh. Get started.
Now, it's going to fail because I haven't installed Cloud Code yet. But this is your first glance at Paperclip. I just hate it's going to be on a failed thing.
Let's go to our inbox. Failed. Let me install Cloud Code on that box real quick.
Just copy and paste the command. We'll run Cloud Code. Get logged in.
And with it installed, we can now go back to Paperclip. And hey, what do you say we click on retry right here? And I'll go back to the dashboard here.
Not sure what's happening. It's running paperclip onboarding. Oh, there's the task.
So let's click on the task. Welcome. I'm Dumbledore, your first agent teammate.
Oh, it's actually gonna walk us through the onboarding. So what you're seeing right here is actually really neat. It thrust us right into a task.
Notice we're on the tasks side or tab of the work and we're here. Look at the commands it ran. It's running a lot of commands here and it finished.
And now we have an inbox item. Let's go see what our inbox is. Okay.
Paperclip onboarding. Click on that. Now, as soon as you set up your company, your CEO's gonna be like, Hey, let's go, let's onboard.
Let's actually make some things happen. It'll do a little questionnaire for you. I don't need that right now.
I know what I'm gonna do. So I'll skip that and just get straight to hiring my employees. First, we're gonna hire our CTO, Ron, our remote Hermes agent.
So really all I have to do is copy this prompt and feed it to my first Hermes agent hire, which would be Ron. Actually, I'm gonna give him a message cuz like he, he knows me. So this point I'm just waiting on Ron to come to the door and he does.
I get a request to approve Ron's hire. And when I click approve, boom, he's hired. He's an employee.
Now notice what Ron says to us. He's like, hey, I'm waiting on board approval. Who's the board?
That's you. Look at us on our company. We're the board of directors.
Look how far we've come. That's where we sit in this paperclip organizational hierarchy. Our agents are the employees.
We're up here. We're calling the shots. Now here, what you're about to see is one of the most powerful parts of paperclip.
And it's how we. talk with our agents, how we interface with them. And that's through tasks.
Now we're about to sign our toilet task, but first I want to just test this and make sure it works. Cause so far we haven't even talked to our agents and our agents haven't talked to each other. We're gonna do both of those right now.
So we'll set up a task and assign it to Dumbledore just to say, Hey, why don't you talk to Ron, test Ron, make sure he works. So we'll set up our task and go. And within moments, watch what happens here.
Boom. Dumbledore assigns a task to run. And that right there is our first moment where we talk to an agent and then an agent talk to another agent.
And they're both working here. Dumbledore's actually waiting. And then when Ron finishes his task, Dumbledore's like, Hey, we tested.
We're good. Isn't that sick? Now some movie magic's about to happen.
I'm going to hire George and Fred. You ready? Set boom.
We've got Fred network engineer and George storage engineer. Look at our org. Look.
Now I'm actually gonna have Dumbledore hire the other employees. So I'll set up a task for each one and just go put Dumbledore to work. And also while we're here, notice the different options we have in a task.
We have different modes. We can set projects. We can do lots of stuff.
Anyways, Dumbledore is working.
All right. We got all the agents hired. Check this out.
My org chart. We got George and Fred on Hermes. Arthur Weasley, which I did not choose that name.
I love that the agent chose that. He's on Codex. Watchdog is on Pi.
Help Desk is on Pi. And by the way, Pi is another harness. And I chose that because it's pretty good with local agents.
I'll show you what model they're running here in a second. Then we got Mad Eye Moody, our security reviewer. I love that choice.
On Codex. And finally, our scanner on Pi. Now we're almost there.
We're about to assign our toilet task and solve all our toilet problems. But first, look at this. You don't even have to come to this paperclip interface.
You can actually talk to your agents and they can do stuff in paperclip. Like I'm asking Ron right now to assign a task to Matt. I Moody through Hermes.
This works because they all have a paperclip skill. Isn't that awesome. Also look at this organization stuff.
We get an org chart. Of course, activity. We have timelines.
Everything's trackable. Everything's traceable. And that's important when you're about to solve a big toilet problem that we're gonna solve right now.
Get your coffee stinking ready. Let's assign that task.
I'll first actually set up a project for our toilet issues. I imagine there might be a lot of tasks that you're gonna go into this project to find, to solve it. So we set up our project and then I'll set up a task and put it in that project.
I got all the details here. Everything I know about the issues so far, I'm gonna give it to Dumbledore and say, go buddy, figure this stuff out for me, please help. And he's off to the races and there he goes.
Now Dumbledore actually does something here I did not expect. He used an experimental feature I enabled called cases, which might be live right now. He creates a case for us.
He's on the case. Oh, things just started happening. He has signed out all the tasks to everyone.
Arthur's got something, George, Fred. That's awesome. Let's click on the main task.
And I think we can go up here and view at the top. We have like options to do things with that task. If I click on these little buttons here, let me zoom in a little bit.
I can pause, subtree, cancel, hide the task, click on properties. I can see a lot more things going on with this. We got blocking tasks, parent tasks, reviewers, approvers, monitors, watchdogs, cases.
Dude, are you seeing the power of this? Now, one thing to note, what's really cool about all this is that if any of these agents have an issue, they run into a block or something, they're not gonna come bothering me with it immediately. They're gonna kick it back to Dumbledore.
It's like, dude, it's not working. Dumbledore's gonna work on it. If he can't solve it, he's gonna kick it back to me, board member.
Check this out, the scanner's being used, which his only job, I think, is just to scan the network. That's cool that that happened that way. Full transparency, I think they're figuring it out, but they're figuring it out.
Things aren't running perfectly. Like this guy, he keeps timing out, I think. But he's scanning, he's finding stuff.
I mean, look at that. He's doing good work. Looks like Fred just finished.
Now, while we're waiting, I want to show you something that I haven't talked about just yet. Over here on the left, under work, you'll also see something called artifacts. What is that?
Let's jump in there. Artifacts is anything that the agents end up making. For example, let's go to map the studio network.
Fred made a report that I can download and take a look at. So let's take a look at the, uh, Mark down file he made here. And I'm gonna have to hide a lot of this, but he did find a lot, a lot of stuff and figure some things out.
I mean, look at this. He's logging into my switches, making things happen. And now all these agents are gonna work together to figure this out for me.
Hopefully. Okay. Something cool is kind of happening right now.
Arthur Weasley built a harness to capture when something happens. And then he assigned a task to Matt. I moody to review the security.
Okay. They finished. And now I have apparently a decision to make.
And also I want to point out that if I go to cases, cases, I guess, closed? Is that how we say it? I don't see the case anymore.
Okay, so let's go to decisions and see what we got. So he's got one question and what he'd like to do next. This is Dumbledore talking after using all the agents to figure stuff out.
When the nest drops, does everyone lose it at the same instant or does each person lose it at their own separate moments? Same instant. And the question two, he said one question.
Liar. Okay, question two, controlled flush test. Nah, they're not all here.
It only happens when they're all here, which is weird, I know. That's another part of the mystery here. Replace the optic.
I'm gonna say other, and I'll say I suspect it's the optic. Let's just see what the report says, and I may replace them today. Then question three, if you approve the flush test.
All right, sending to Dumbledore now. Now, while he's thinking, I wanna go look at the timeline, just see how this looks. A lot more stuff going on today.
Let's zoom in. Look at all that. Oh, I love it.
I love the tracking. Let's go look at the artifacts. I think some cool things were made while we weren't looking.
Artifacts. Ooh, an audio, or not an audio, a studio network map. Let's take a look.
Wait, where'd it go? How do I view it? Okay, it took me to the task.
That's annoying. Oh, open it? Oh, there we go.
I'm probably gonna have to hide this from you, but... This is very accurate. So leading fault signature is that my SFP on my Mac logged six link up down events today and storage shows no matching events.
So it's not storage. Okay, stuff's happening now. He has signed a lot more things out.
He's having Fred look at spanning tree and Mac table flushes, having Ron correlate some stuff and checking all NAS clients so they all disconnect at the same time. Oh my gosh, this is so cool. Okay, I got another decision to make here.
They're still working though. They're still going. I don't want to replace the optic today.
I think Ron's investigating thermals right now, so I'm going to replace it, but hold until that report. Skip the flush test. WAN failovers happened yesterday.
That's news to me. What the heck? Yeah, open a ticket for it.
This is already paying dividends. I love this. Now, while Dumbledore's still working, I do want to show you one thing.
Routines are actually pretty sick. Let's create one right now. You can think of a routine like a scheduled task, right?
It's gonna run on a schedule that you set and we can do a ton of things with this. It's actually pretty powerful. Let's go ahead and create one.
Got the option up here at the top right. Create routine. And I'll do something basic just to illustrate what we can do.
Check storage. And I'm gonna voice dictate this because I'm lazy. Every day, check the storage on the Ceph server and check for any health issues.
I want you to report back the amount of storage we have left. And if there's anything that's low, you need to create a decision for Chuck. Okay.
I would make that obviously a lot better if I had time, but let's make it Fred responsible. No project. Uh, we have some advanced stuff.
Not worried about that. Ooh, Dumbledore finished. He already finished the failover stuff.
Anyways, creating the routine. Then we have triggers to make this thing happen. Oh, we also have.
Secrets and variables. We can't do that. Can't do it right now.
We gotta focus. So our triggers, I could add a schedule. It's like every day at 10 a .m.
Add trigger. Boom. It's active.
It's saved. If I go to routines, I see it sitting right there. You can also run it at any time.
By the way, this is how I run my standups. Run now. Run the routine.
And Fred's running. Dang, Fred already finished the storage check. I got my decision in here.
SSD tier capacity is low. Only 29 terabytes available, which sounds like a lot. Cause it is a lot.
Why is it so low? I can review and plan SSD capacity relief. I'll defer for now.
But how cool was that? Now did the network drop issue finish? I think it did.
So this is like Dumbledore's last answer. Final report. It's right here.
Okay. So it's not the NAS, not the router, not spanning tree, blah, blah. And importantly, it's not one bad port dragging the others down.
Okay. This is getting interesting. NAS never went down.
Now they did the router, micro tick, unify switch. Nothing was changed on any machine. I mean, they went deep.
Look at this. So their honest report is that they can't name the trigger. So not established that it's the toilet.
A flush does fit everything we see though, but so would a door, HVAC compressor, cycling, or someone's chair hitting the cabinet near the patch field. But I do want to point out, it did say, a shared external physical event drops four specific fiber runs on flu powder, which is my switch, two seconds, 36 times in the retain week.
But this is sick that we got all this data and you know what? It's all stored here than the company. It's all like artifacts.
It's all findable and traceable by every agent. And right now Ron's working on the other issue, which is cool. Okay.
Some time has passed and the issues still persisted, but now I have a real verdict. I was tempted to let this issue just kind of go, but no, it got annoying. Like literally today I was in a meeting.
Boom. my connection goes down. I stop and I walk out and I go, did someone just flush the toilet?
And Mike was like, yeah, it was me. It's still here. I'm not crazy.
So I had them dig deeper. And this is the official verdict. It's not the toilet.
I don't believe him though. And it's not a mystery. Then what is it?
Four links on flu powder had been broken for 15 months. Flu powder is my micro tech switch. We're talking about anything that disturbs the building knocks them over.
Nothing else in that switch ever notices. So my theory, when someone flushes the toilet and specifically Mike, cause Mike, you always know Mike's around because he's always banging stuff around. He's he's Mike.
I think when he uses the bathroom, he flushes it very forcefully. He opens the door very forcefully. And that motion causes something in the adjacent room where my server room is to jostle things.
But that alone is like, okay, well, how does that, why just those four links? It's two faults stacked on top of each other. Ports three, five, and 13 and 19 have no margin.
They lose sync 56 times a day. Port five is logged 20 ,052 link downs since the switch is booted. A comparable port has only done 123.
I think they're clocking at like 40 something odd disconnections a day. And dude, I know my editors are cursed, but they're not flushing the toilet 40 times a day. And something occupancy linked, meaning like people being there.
knocks two to four of them down in the same second, seven times every working day between these hours. So only four of the ports are close enough to the edge to fall over. So what is it?
What do I do? It comes down to the transceivers. So I have these SFPs.
They are these third -party SFPs I bought off Amazon, fairly cheap. I bought them in different batches over time as I've added more employees. And apparently the effective ports are all in one of the batches.
Whatever this is. This batch has a 50 % in service failure rate and every port we've chased is in it. So the answer is this.
I actually bought some brand new SFPs. These are not third party. If you look at them here, they are MicroTik branded SFPs.
I only bought three. Actually, I can already tell these feel good. So the solution, the real solution is replace the modules and patch these guys.
Different vendor, different batch. i went with the official vendor let's go replace them right now all right let's go hey you may get disconnected for a moment while i replace some stuff in there in the toilet room yeah okay Wasn't this whole thing just crazy?
Like we set up a company, we had employees, they were doing work, they were talking to each other, commenting on each other's tasks, and they were all simultaneously working to solve this problem. Gosh, who read this? I still think it's the toilet.
The correlation's still so strong. It might be surrounding, the events surrounding the toilet thing, but I still think it's funnier. Too coincidental for it not to be the toilet flushes.
It's a better story anyway. We'll see if this solves it. I'm going to let it run for a week or two before I say done.
This is a verdict. You'll have to subscribe to find out if that is indeed what solved my toilet problems. Could we go deeper and try to do some environmental testing around it?
Try to, I don't know, do some crazy stuff like electromagnetic whatever. Sure. I don't have time for that.
And why do these all come in different packaging? Maybe I'm screwed, I don't know. Now, I know you probably thought I forgot about this, but I didn't.
During that Flare ad that you watched in its entirety, right? I said I was gonna connect Flare to Paperclip, and I'm gonna do it in kind of a fun way. So first, I'm gonna have Dumbledore hire somebody, because now I'm so lazy.
Hire this dude, Dumbledore. Have him use Cloud Code Opus 5. So check this out.
I'm gonna go to Settings, go to Secrets, create a new secret. Call it flare API key, paste it and create it. And I'll give it access to my agent that I haven't created yet.
And then I'll finish creating our agent that we were creating before with that task. Dumbledore's got this go, go, go. Oh snap.
I was not ready for this. He already hired Filch. He's going.
I don't think Filch has access to my secret though. Let's get back to the secrets. Filch.
Let me give you a secret, dude. Here you go Filch. Add.
Okay. Filch has it now. That's actually really sick.
I just gave it to you, buddy. He said he failed. I gave it to you.
It is here. You don't have to worry about this. I'm going to verify if it's actually in his secrets.
It is. It shows up right there. That's awesome.
He also has a flare skill that I created without you looking and we'll see what he comes up with. Okay. I got a decision.
Oh, he's like, I'm missing it. Flare tenant ID. I didn't set that.
Yeah, that was my bad. I didn't set that, but they already knew it. And that's so cool.
New secret. Flare tenant ID. Create.
Give it to Filch. There you go, Filch. Go to my decisions.
Both set. Run the sweep. Boom.
This is actually pretty fun. Credentials are present. Reading the skill.
All right, we'll wait until he's done. Okay, I didn't ask for this, but Filch found stuff, and he's got Ron doing some forced password resets on things? He's got Mad -Eye Moody checking some stuff?
Dude, what did he find? Short answer is he found a lot. And he assigned lots of tasks.
You get mad -eyed judging. I don't think Ron has the ability to do that. So let me show you this actually real quick.
If you're like, oh, they shouldn't be doing this task. This is crazy. You can go over here and like pause the work.
So like, you don't need to do this. Pause it. Also, one more thing, just one more thing.
If I go to my org option here, notice at the top, I can import or export my organization. So everything we've done here. everything, every agent, every routine and task.
Ah, attachment. We can export that. This is ours.
This is our company. We can export and import somewhere else. Oh man, I forgot to show you how, what I do, um, my standup every day.
So I do have another paperclip I've been running for a while. Oh, and Oh, the standup is running right now because it's five o 'clock. Uh, literally happened.
So all my agents are running their standup. It's actually a routine and each agent will go through and talk about what they did that day. They will also talk to each other.
If there's anything that's not clear about what they heard from other people. You know what? I forgot to mention.
I have an agent named Luna. She's my thumbnail agent. I have 10 agents in this.
I'm gonna show you the result of the standup. And it's actually been a pretty light day because it's a holiday and only I've been here working. Ah, see, here's what I was waiting for right here.
There's a Q &A between some of the agents because they had questions about what each other were doing. I'll check that out. Here's a question that Taylor had for Hermione.
Look at that. Taylor's trying to make sure that we have backups for our footage because we have restream recordings being expired. How sick is that?
Oh, I love that example. Here we go. I gotta find that brief.
Yeah, and they did all the work. It's super cool. Let me know what you think of Paperclip in the comments below.
Are you gonna use it? You can start that repo because you should, because Dota is amazing. And please let me know what other things you want me to try, what other harnesses or AI tools out there that you think are interesting.
Maybe you have one. Comment below. That's it.
That's all I got. I'll catch you guys next time. Hey, you're still here.
At the end of my videos, I like to pray for you, my audience. I put it at the end because I know prayer isn't everyone's thing. But if you made it till the end, that's not an accident.
Chill out with me for a little bit and let's Let's pray. Let's go with it.
One, two, three, pray. God, I ask in your name that you bless the person on the other side of this camera, that you would help them right now with any stress or anxiety they have. I know for me today, I had a ton of anxiety.
And just praying through that, giving it to you, Lord, helped me so much. So God, I pray peace and stillness over.
the person on the other side of this camera. Peace and stillness over you right now. Whatever is bothering them, whatever is controlling them, whatever is overwhelming them, let it become small in their mind.
Let it become manageable. Let them just sit down with it and write it down and turn it into smaller pieces that won't overwhelm them. And just, Father, give them peace beyond all understanding.
I thank you for who they are and what they're about. that they're passionate for IT and just continue to just make them excited for this. Light them up, God.
Give them opportunity. I ask this in your name, Jesus. Amen.
All right, that's it. Thank you for letting me do that. That's all I got.
I'll catch you guys next time.
The Hook
The bait, then the rug-pull.
Every AI agent lives in its own chat window, answering to nobody but the person who opened it. This video hires eight of them, across four different AI companies' tools, into a single org chart with a CEO, and then asks whether a company of AI employees can actually solve a problem a solo agent couldn't.
Named ideas worth stealing.
08:48model
Company hierarchy for AI agents
- Board (human)
- CEO agent
- Managers (e.g. CTO)
- Direct-report agents
Paperclip organizes agents into a literal company: a human board member sits above a CEO agent, who manages department leads, who manage individual worker agents. Blockers escalate up the chain instead of interrupting the human directly.
Steal forany setup running more than 2-3 agents that currently have no reporting structure or shared task queue
15:51concept
Human approval gate pattern
- Agent surfaces a decision, not just an update
- Multiple concrete options are presented, not a blank prompt
- Human picks one before the agent proceeds
Instead of asking an open-ended question, agents that hit a fork in the investigation present the human with named, scoped options (run a test, replace a part, audit a path) and wait for a pick.
Steal forany autonomous pipeline where a wrong irreversible action (spend, physical part swap, production change) needs a human checkpoint
How they asked for the click.
VERBAL ASK
27:00next-video
“Let me know what you think of Paperclip in the comments below. Are you gonna use it? ... Please let me know what other things you want me to try, what other harnesses or AI tools out there that you think are interesting.”
Soft CTA framed as a question rather than a hard ask, invites viewers to name other harnesses/tools for future videos, plus a sponsor CTA to try Flare earlier in the video.
MENTIONED ON CAMERA
Visual structure at a glance.
