High-Level Agentic Engineering with David Ondrej

0:00 My name is David Andre and over the past four years I've spent over 2,500 hours coding with AI and I've used that to build three different AI startups. I've also personally interviewed some of the

0:10 most productive people when it comes to agent engineering such as the CEO of Versell, CTO of Superbase, CTO of Fireworks, and Matt PCO. And let me tell you, when you know when you crack it how

0:22 to use AI to write code, it really does feel like cheating. So, in this video, I'm going to show you the best models and harnesses right now, and which subscriptions to buy, how I review code

0:31 so I can ship straight to production with confidence, when to use sub agents and when to use cloud agents, and the loops that do a big part of my daily work for me. And when I did this video 2

0:41 months ago, it nearly got 200,000 views. And let me tell you, a lot has changed since the last two months. So, first of all, let's talk about what are the best models right now. Harnesses and

0:51 subscriptions. This is the main thing. How do you get the most value? So let me jump into cloudroom. What I use the most above all other model right now is Opus 5.5. I think it's the best model in the

1:00 world but also it's very fast. It's obviously the top model from Enthropic and maybe Fable 5.5 is around the corner but still Opus is incredible release. I would say it's one of the best releases

1:12 from the last two years. So either I use it inside of cloth code but right now I used up both of my subscriptions. I might need to get a third one or I use it inside of some other harness such as

1:23 cursor CLI. Right. So, Opus 5.5 as the model. Best model from OpenAI. It's between Astra and Soul, but I think Astra has that big model smell. So, I would say this is the best model from

1:32 OpenAI right now, but they really are falling behind and they need to release a better model badly. Okay, so those are the two best models in the world. What about other models, right? Because

1:42 there's other companies, Google, Meta, Kimmy. The thing is they are good and you should still use them. I will go later in the video how I use them, but they're not the best. They're not at the

1:53 frontier, true frontier. That's where Opus and Astra are right now. Now, when it comes to the harness, obviously I use multiple different harnesses when it comes to Pi, Cloth Code, Hermes, Open

2:02 Code, but the harness is starting to become less and less important. And in fact, last week we were at a event at Enthropics office and I spoke to two different engineers at Enthropic and

2:12 they told me kind of what they see right as it comes to like the next next generation of models. And basically the main lesson was that the harness is going to matter less and less. So

2:20 whether you run Opus or Fable in close code harness or in Devvin or in Droid if you give a powerful model access to terminal access to the internet it doesn't really matter what

2:32 harness you run it in. Um so again take it from those developers at entropic. When it comes to subscriptions this is where you can get a lot of value. Right now the best value is claw subscription

2:42 both 100 and 200 much better value than codex. Second most value is OpenI. Um you know the CHI GBT pro subscription but these guys they keep changing it. Like

2:53 literally today they like uh returned it. They wanted to half it. They they doubled it again. They released the $500 plan. I don't know. Open AAI is very inconsistent over the last month. They

3:04 lost a lot of their reputation. But if I only had to choose one, it would be the Claude Max 200 subscription and it's not even closed right now. So when it comes to other subscriptions, right, if you

3:14 don't have much money, Open Code Go is a great $10 subscription. If you like Grogbot and some of the products from Cursor and SpaceX AI, then obviously their subscription is fine. Google is a

3:25 bit of a joke. Although Argon is promising and the new model Carbon, which is again not released, seems to be even more promising and actually probably the best in the world. Again,

3:36 it's not released, so it's kind of a pointless release. Now, in terms of the app, you know, the ADE, agent development environment, I only use Cloud Room nowadays. So, Cloud Room is

3:45 an AD I've built. It's completely free and open source. You can get it at cloudroom.dev. I'm going to also going to link it below the video. In fact, it's the only AD in the world. The only

3:54 AI coding app in the world that gives you cloud agents on the free plan. So, you can run any model, whether it's Opus, whether it's Astra, whether it's Grock, whether it's Muse in Cloud

4:05 Agents, right? No matter what's your favorite model, your favorite harness, your existing subscriptions, you can use all of them both on local and cloud. And again, it's completely free and um fully

4:15 open source. So go to cloud.dev and get it. And if you like the project, I would appreciate if you start it on GitHub. Now, let's talk about how to be a great agentic engineer because there's a lot

4:25 of u questions like should I learn how to code? Will this even exist in the future? Prompt engineering, how important are prompts? a lot of noise, a lot of nonsense from people who have

4:35 never built anything with AI or from people who are like, you know, come from the traditional coding background. They've been in software for 15 years and they still have a lot of these

4:45 habits from the past. So, what skills should you learn? Is it still learn to code technical depth versus breath? So, first of all, you shouldn't try to like learn syntax and memorize specific like

4:57 you know words and you know, Python and Typescript. You should learn the difference between JavaScript and TypeScript, right? Like what is a type language? You should know like why Rust

5:06 is fast or how Rust is memory safe. You should learn these kind of high level concepts. But looking at a piece of code and knowing, okay, this is C++ or this is Rust. This is not really valuable,

5:17 right? Sure, it does. It's cool that you know it, but it's not really valuable. What matters is thinking like an architect basically. And there's a great video from Mark Richards on this. And

5:26 you need to kind of have an awareness of different technologies, right? So you need to know like okay here the front end is kind of slow and may maybe like you know this like if you switch to a

5:35 different page if I can go to change log hm why is that instant maybe it's like pre-loading right maybe it's caching maybe it's optimistic loading you need to know these kind of high level terms

5:44 as if you were a software architect that's really the best skill set the best way to describe the skill set it's having larger breath less depth in each of them and larger breath right so for

5:55 example if I show you something like this everyone who's building software should understand these concepts and know something of them, right? Trade-offs, caching, observability,

6:04 blast radius, coupling and cohesion. You should kind of understand what that means and you should have some awareness of this. Doesn't mean you need to be an expert in them. Doesn't mean you need to

6:15 know how to write the function from scratch doing that because you can just use Opus 5.5 to do it. You should know these concepts exist. So that's the best way to explain it. And again, just feel

6:24 free to like look into any of these concepts if you don't understand some of them. just spend 30 minutes learning them. Like the best way to learn is with AI as well. So actually using AI to

6:33 upskill yourself is one of the highest predictors of success I've seen. And by the way, later in the video I'll show you how you can actually build this into a self-improving piece of software. What

6:42 that means is as you have more users and as your software grows, it literally is self-improving itself as autonomous self-improve self-improving software factory. So we're going to touch on this

6:51 later in the video. Now let's talk about reviewing code and my own methodology of how I am able to ship hundreds of changes per day into production. And this is not a exaggeration. If you go to

7:01 my GitHub, you can really see the acceleration right here. I wasn't really building anything. Then I started building um again earlier this year in like late March, April. But you can see

7:11 over time the acceleration kicks off and really in uh September and October the more I started embracing cloud agents my productivity just skyrocketed and now I'm like shipping 100 to 200

7:24 contributions to production every single day right and when I visualized it earlier on the different chart it looks like this because sometimes on this GitHub chart it cannot be it's not that

7:35 clear but then I told you opus is great visualization so I told it to scrape the actual data and to visualize it on this different chart. And now I'm I'm doing 700 commits to production per week. And

7:46 it really is growing exponentially for me. So it's important to also quantify your productivity to make sure that you know if you're um getting more productive with AI or if you're just

7:57 wasting time and kind of building internal tooling, which a lot of people are just doing a lot of that. So reviewing code, let's talk about this. And in fact, this might be one of the

8:07 most valuable parts of this video because this is actually my workflow. Now, personally, I could not even create this, right? Cuz like if I if you asked me to describe it, I would not go into

8:16 such detail. What I did instead is I asked Opus to analyze all my sessions inside of Cloud Room. In fact, if I say how many um open threads I have inside of Cloud Room right now, it can use the

8:28 room CLI to check how many open threads I have across all the different repos and projects because I think it's like 500 plus or something. Um yesterday it got over 800 and I closed a bunch of

8:38 them. Yeah, 550. So, you know, over the past few weeks is literally thousands. How do you know if it's good code? How do you know if it's quality? How do you know if there's no bugs or critical

8:47 issues or errors or security vulnerabilities, right? You need to know how to review code. So, this is becoming more and more important. And here's exactly how I do it. Like, this is

8:55 literally how I work with AI after thousands of hours. So, task in cloud room thread, right? Very simple separation. By the way, I need to clarify it because some people are

9:05 mixing stuff. When you have a new thread, keep it as a single task, right? Single thing it's doing. Uh, this guy was fixing like my OBS was literally crashing as I was about to record. So,

9:15 this threat was investigating that, right? Here I was preparing for this video. Uh, some other one was like cleaning up my storage. This one was specifically for the cloud room threats,

9:25 right? Don't mix work. Just like u you should have a clear task or clear Jira ticket on a single thing. That's how you should work with agent sessions. Don't try to mix and match. If there's

9:35 something else, you can just do a fork, right? And fork it into a new session or something like that. But each thread should be one clear task, one clear feature, one clear bug fix. So let me

9:45 start there. So a task is in a cloud room thread with the first thing is is if it's a bug, right? Because sometimes it's a feature and or like a research, right? So if it's a bug, then the agent

9:57 tries to reproduce it end to end like a real user. Don't try to fix bugs by with a vague descriptions, vague description. You can have the agent reproduce it. You can have the agent run a test, run a

10:08 simulation, run an experiment to try to understand the bug more and reproduce it. Then Opus 5.5 builds it, the fix, and then it goes into some test type check and screenshots of the real app of

10:20 how it will look like when it's fixed, when it's a UI change. This is also very OP. Um, asking for screenshots, right? So, I can say like if we were to redesign our app in a more like

10:30 traditional software, let's say it's like a Windows 7 type of app. Look at the attached image. If you were to redesign that in a Windows 7 style, what would that look like? Show me free

10:40 potential image renders. And by the way, we do have Windows support. Now, a lot of you guys couldn't use Cloud Room because we only had uh Mac OS and Linux support. And we do have

10:50 Windows support officially as of yesterday. So again, cloudroom.dev. If you're a Windows user, download it and give it a shot. So, Opus does the fix, runs the test, and if there's some UI

10:59 change, it shows me screenshots. Now, if it's not a bug, it asks, is it big or risky? If so, I have multiple skills on this. And by the way, I'm going to also link my skills repo. This actually my

11:10 most popular GitHub repo where you can get all of my personal agent skills. Again, fully open source and completely free such as 4x4. So, it's like run 4x4. it um pulls 16 decisions quickly out of

11:23 me in different formats. So I can actually have to explain explain how 4x4 skill works. Be very concise. If you see me typing, you'll see that a lot of these u common prompts I have on raycast

11:36 snippets which I guess a bonus tip just no matter if it's the Mac OS snippets like the text replacements which is like built into Mac OS text replacements. Boom. You can go here and you can see um

11:49 your different text replacements in Mac OS. But even better is like raycast snippets which is just faster. So all the things that like I'm I'm typing in um frequently you know often I just have

11:58 them on raycast snippets. I can work so much faster. Like imagine if I had to like retype all of this stuff every single time by hand. There's no way. If it's something more complex than a

12:07 single prompt, then it's a skill, right? Such as the 4x4 skill which is explained right here. And this asks me four sets of four questions. first about product then software design then practical and

12:17 then edge cases and to be fair I don't run it on small changes only on big changes I have this skill risky changes which is automatically triggered it's kind of funny because when I saw this I

12:26 almost forgot about the skill but then I realized the agents are running it all the time when they're doing risky changes right so I can say explain risky changes

12:37 this gets triggered is a safety check for changes where a mistake would hurt customers why it exists one filter passed every on live data it turned off 99% of the feature. So yeah, usually my

12:50 skills exist from a expensive lesson, right? So if it's a big or risky change, it runs these uh skills. If not, Opus just builds it. A lot of the times it's kind of obvious it's a small thing and

13:02 then it goes to the tests and type check. By the way, when it comes to tests, I'm a very contrarian on test. They are usually overhyped. In fact, you almost need to tell your agents to not

13:13 write tests unless it's truly necessary because over 90% of tests that the agent writes are complete slope. It just blows your app. It slows you down and it doesn't test anything. They just write

13:24 the test for the sake of passing. Sure, they still have um purpose. If you see your agents doing the same type of mistake or the same class of errors, then tell them to implement one or two

13:34 strategic tests to prevent this from happening in the future. But unless your your app is failing and unless there's a problem, don't bloat it with hundreds of thousands of lines of code of tests.

13:45 Okay? The models are good enough and the models are only getting better. And and the models by themselves write a lot of tests. So to balance it out and to hit the right amount of tests, you need to

13:54 tell them to stop writing them and they will still write some of them. And that's usually the right amount. Okay. So that's the build part, right? Then we go to the review part of uh when

14:04 something happens. So again, how big is it? If it's normal, I just run GPT review because again, I'm running with Opus. And this is a critical thing to understand. You need to use frontier

14:14 models. A lot of the things I'm saying guys doesn't make sense for you if you're using cheap models. If you're like only using Haiku or only using Luna and trying to cost optimize everything,

14:24 this will not make sense. You'll be like, why is David saying about shipping straight to production hundreds of commits per day when my agents are failing and they're they're doing

14:33 mistakes? I'm using only Frontier models like it's completely different if you're using cheap cheap model or like a 27B locally hosted model. A lot of this is not going to be applicable for you. I'm

14:45 using Opus 5.5 extra high or max on everything and u when I'm reviewing it, I have GPT review. So I can just go here explain GPD review and this reviews it with the other frontier model. So

14:58 everything you do if it's a medium or large change you should review it with the other frontier model. So if you have enthropic right so let's say you're a enthropic fan and you're using opus or

15:10 fable well guess what you should run GPT review right you should use openi model the frontier model from openi so right now it's arstra or gbt 6.1 soul with extra high or max to review whatever

15:24 happened through opus or through fable if you are a GPT fan and you prefer open models cool just make sure to review the the big and risky changes with the latest frontier an entropic

15:36 models such as Opus 5.5. If you just do this one simple thing, you will minimize the number of like serious bugs and and breaking changes going and the number of slop. You'll just minimize it by like

15:47 90%. I'm telling you guys, review big and risky changes with the other frontier model and you're good to go. If it's a huge change, okay, like a huge risky change, I have this skill which I

16:01 don't use it that often because it's slow. Total review. And this really throws all the tokens at it, right? You use Fable, Opus, Astra, Solo, Muse, and it's uh reviews those changes with all

16:13 of the Frontier models from all of the labs and then gives me a ddupl list, dduplicated list of those most serious findings at the end. Again, I only use it like few times a week when

16:26 I'm really not sure and when it's actually a huge change. Most of the time I build with Opus and then I just run GPT review. If it finds a lot of stuff, I fix it and I might run it a second

16:35 time. Another bad mistake I see often is running this multiple times in a row, right? Like open review, GP review, GP review, GP review. This is slop because these models when you tell them find

16:46 bugs, even when there are no serious bugs, they're going to find something. They're going to point out something. they're going to hallucinate it, invent it because that was the task you gave

16:54 them, right? Part of this is every issue gets rated 1 through 10, especially with Astra. It can uh spot like edge cases, which might be serious, but might never happen. So, it rates it on two axes, how

17:08 often and how bad it is, right? It might be very bad issue, but it's like extremely rare scenario that will never happen. And if you don't have much time, you would skip that, right? The problem

17:19 is when it's a serious bad problem and it could happen often, right? And again, you can use your best judgment. If it's a small thing, you can decide to, you know, if it finds six issues, you can

17:30 decide to just fix the most serious three serious three issues. Usually, the best way to do it is just send it back to Opus, right? So, let's say you have Astra, it found like five problems. You

17:40 ask Opus, do you agree that all five of these should be fixed or are some of them overthinking? And usually what will happen is it will tell you, I agree with four of them. The last one is is really

17:51 never going to happen. So you just fix four of these and you push to production. It really is as simple as that. So let's say there's an issue, right? And again, I ask Opus, which is

18:01 my, you know, driver, my go-to. Is this a real mustfix issue or overthinking? This is a great prompt. It caches a lot of things. If it's must fix, then I just approve it and it fixes it. If it's

18:11 unsure, u you can run a second opinion with a different review. So, Opus review, news review, um, Astra review, fable review, whatever. And again, this is u optional

18:25 re-reviewing the fix. Most of the times it's not needed and it slows you down. It also depends like if you have thousands of paying customers, sure, you need to be more careful. If you are just

18:34 starting out, which most of you don't have a massive codebase with, you know, enterprise contracts and all that, you need to be building fast. You cannot be re-reing everything 10 times. you need

18:44 to build fast and break things. Um and if you are not sure then you review. Okay. So once there's nothing left to fix the last section is to ship right. So ship to production pull rebates main.

18:55 There is a CI obviously you can have some checks and um tests there. Side effects security loops. Okay. So, I'm going to touch on my agent loops later in the video because a lot of the

19:05 problems I catch them myself before they would get to customers because I have agent loops kind of testing recent changes and what they might break. The the worst part is like if you're

19:15 shipping fast and even if you're using reviews, they might be side effects to different parts of the codebase that you did not expect, which again having like one agent loop dedicated to that catches

19:27 95% of these. So yeah, this is my kind of a process, my logic of how I work day-to-day. Um, probably will change a lot in a month, but hey, I'll make another part of this video. And by the

19:38 way, a great article I've seen from Jacob. I might even make a full video on this is this kind of the mindset is the closest to I've seen most people even here in San Francisco right now I'm NSF

19:51 because you know to meeting investors for cloud room and all that stuff most startup founders I've met still do not fully embrace AI right and the the reason why this matters is I'm

20:01 going to explain this to you with a via very visual way is if you have 100% of your work let's say you do 50% manually now and you embrace AI, you think you're embracing AI and you do only 50% by

20:19 agents, right? So let's say before if you worked eight hours a day. Now after embracing agents, maybe you only work four hours a day. Okay? And the other four hours that your

20:33 agents are handling could be done in like 10 minutes, 20 minutes. So instead of 8 hours something would take you 4 hours 10 minutes right because again before it would be four and four now you

20:46 only do the four manually and the other you aggressively use agents it's 4 hours 10 but you see the issue the productivity gains are not that huge because you have the time or like you

20:57 know 45% reduction whatever of the time it takes you the only gains the only time where you really see insane improvements is where you go 100%. When AI writes 100% of your code, because

21:09 that's where you can move 10, 20x, 50x, 100x faster, when agents are writing everything, every single line of code and reviewing every single line of code. Because if you're still writing code

21:20 manually and like you think you're one of the most locked in engineers, you think you're embracing AI because half of your stuff, half of your issues or tasks are handled by agents. You are

21:32 still the bottleneck. I I really need to drill this into your head, guys. You are still the slow part. You are still the slowest moving cog in the machine. Even when you use AI for 50% 60% of your

21:45 work, you need to take the red pill like in the matrix. And you need to use AI for 100% of your work to really see the productivity gains. So the TLDDR of the mindset is like agents fix every error

21:57 hit by a user. Same for me. And verify their own work with no human in the loop. This is the part I'm still working on with agent loops and trying to remove myself as much as possible. And really,

22:06 this is the part we will see in the last quarter of 2026 and 2027 is like the people who will just pull ahead of everyone else. It's because they'll run their own software factories. They'll

22:16 run a series of agent loops that run without human in the loop. And this is an important mindset. I don't really care if a agents break production. What matters is that they fix it themselves

22:27 and fix it fast. Right? So this is how to review code. And this is my kind of a process for shipping. Now let's talk about sub aents because this is a topic I have a lot of opinions on. So I think

22:38 uh Max in this tweet said it well that sub agent orchestration is basically astrology for vibe coders. A lot of people are like obsessed with oh I have this many sub agents or that many sub

22:50 agents when in reality you could have done it better with a single agent that is a top level agent. Right? There's a lot of questions like when

23:00 should I use sub agents? Isn't this just wasting my limit? Which model should I use sub aents? All of that, right? And I thought about this a lot and the best way to explain it is with this diagram.

23:14 When to use sub agents is where the work is clear and parallel and can be chunked into predictable sessions. Right? So, for example, if the main agent needs to analyze 2,000 different files, maybe you

23:26 run a law firm and you have 2,000 different documents or maybe you want to analyze your own agent sessions, right? And there's thousands of them and it's the same predictable process. You just

23:36 need to categorize. You just need to read it. Maybe that you downloaded a bunch of PDFs. This is the ideal work for sub agents. You use a main agent, you know, with a smart model like Opus

23:46 5.5 and then you run small sub agents like Haiku 5.5 on those specific chunks. You run them all in parallel. You have the main agent start 10 or 20 sub agents and give him the results. This is only

24:02 good when the work is clear, predictable and doesn't have any like unknowns, any question marks, anything crazy. The issue is most of the work is not like this. I don't know, maybe for you

24:15 it is, depends what you do. But if you develop software and if you're building new stuff, most of the work includes surprises, dead ends, you know, unexpected changes, includes reading a

24:26 lot of stuff and figuring out what was the bug, figuring out how do we implement this new feature in the best way possible. And for that, you really really do not want to use subsations,

24:37 right? So for me most of the work doesn't include any sub engines. I want to use a frontier model either Opus 5.5 or Astra or GPD6 all and I wanted that model to figure it out. Right? So a lot

24:52 of people when they LAR how many sub agents they're using and you know use Grog bot to delegate to some other Grog bots and blah blah blah. If you're doing something innovative, something

25:02 new, you really don't want a small model to like try to fix a small bug and then it realizes, oh, this actually coupled with a different part of the codebase and this actually touches billing. That

25:13 is a total disaster. Like total disaster. And you really don't want a small model touching some parts of the codebase. When you're fixing a bug, when you're implementing a new feature, doing

25:24 a major refactor, you want to be using the smartest model possible because anytime there's a surprise, anytime there is opportunity for optimization or, you know, anytime there's like a

25:34 duplicated functions or patterns that need to be fixed or or antiatterns in in software, you want to have the smartest model with maximum reasoning effort looking at it. The only situation really

25:45 where subisions make sense is predictable work where it can be chunked into clear segments into clear buckets that are all the same and where you know exactly what the input data is like and

25:58 exactly what the process will look like and exactly what the output format should look like which to be honest most tasks are really not like that. So this is how you should decide if you should

26:08 do sub agents or not. If you're not sure the answer is no. Most of the time the answer is no and you should just use a single powerful agent and in fact if you look the way I work I just have most top

26:19 level agents running frontier model. I don't try to like split stuff into sub agents to like get it done faster. I never use these like ultra code or ultra mode from uh codeex. You have no idea

26:30 what's happening there, right? Like you are literally wasting your tokens. you're like paying five times more in some some cases for like you know it's getting done 2x faster and you know

26:41 given how valuable these subscriptions are you really don't want to be spending your limit five times faster to be getting something done like you know 1.5x or 2x faster. So sub aents

26:52 massively overhyped very few situations where it actually makes sense. It needs to be like clear chunks of a parallel work that just makes sense. Now let's talk about cloud versus local. Right?

27:02 Cloud regions are a big trend that's happening this year and really didn't exist before the start of this year. This is the famous chart from cursor. At the start of the year only about 1 in 10

27:13 pull requests that made it into production came from cloud agents. In the middle of the year it was around 50% and now it's growing towards 60 70% as we're approaching the rest of the year.

27:24 Now my own growth was similar although I started adopting cloud agents later for different reasons right so before I built cloud room every single cloud agent solution was closed source right

27:37 that's why cloud room is literally the world's first open source cloud agents and the issue with that is that you you lock yourself into something that you cannot inspect you cannot self-host you

27:47 cannot fork it you don't know what's happening you cannot like inspect the GitHub repo right if you go to cloud room you can literally run your own agents on the GitHub repo to check if

27:56 it's safe to check how it's architectured to see if there's no optimizations and if there's some issue you can do you can file a GitHub issue or do a pull request you cannot do that

28:04 with Codex cloud or cursor cloud or am cloud or Devon cloud there's so many closed cloud solutions that their own their entire business model is to lock you in maximum ecosystem lock in they

28:18 force you to use their own models they force you to use their own subscriptions it's insane Everybody is paying big money for Claude Max or CHBD Pro, right? I have multiple

28:29 $200 a month subscriptions. Everyone on my team has multiple $200 a month subscriptions, whether it's from Claude, whether it's from Cursor or Codex. You want to use those tokens, right? And you

28:40 don't want to use some um cloud agents which prevent you from using these. Well, guess what? We support all of the harnesses in Cloud Room. And actually um this was the main innovation that I

28:50 invented letting you run any harness as cloud agent. So that's why I'm saying I was very skeptical of cloud agents for the first 60% of the year because anytime I tried them it was like it was

29:03 tempting but I understood I don't want to build my whole business inside of this closed source company not knowing what's happening and maybe they you know like with cursor cursor was great until

29:14 openAI removed their models and when Astra was the best model in the world for the three weeks you couldn't use it inside of cursor right so imagine if you built all of your cloud environments all

29:24 your API keys, everything on the cursor cloud agents and suddenly Astra when it was the best model in the world is not available and of course you cannot use your cloud subscription or or your open

29:36 AI subscription to power your cursor cloud agents that is the thing you guys have to understand now what let me explain the concept of cloud agents and why it matters some of you might be

29:44 skeptical the thing is agents take up resources they need CPU they need RAM they need disk if you run agents in the primary check out. You'll run into issues of them conflicting,

29:57 right? They might start changing the same files. They might fight against each other. The solution is work trees. It's a decent solution. Um, it's what I used for a couple of weeks. The problem

30:07 is each worksheet cop copies the entire primary checkout, which can be gigabytes as your project grows, right? So, it fills up your disk very fast. The second issue is still with worksheets, your

30:20 agents are running on the same machine. If your agent decides to run a heavy test suite or do some rendering, do some computational work, analysis of the disk, chunking something, anything that

30:31 re requires a lot of memory, a lot of CPU, a lot of GPU, or if your agents decide to click around and use computer use, right? Maybe like move stuff around, open new browser tabs, do some

30:42 pop-ups with some apps, it's extremely annoying. You're working on the same computer, there's pop-ups happening, there's a, you know, it starts to lagging, you're starting to hear the

30:49 fans. This is the real limit of running everything local. Sure, maybe it doesn't affect you if you're new to AI and if you're running two free agents, but if you want to go to the next level and run

30:59 10, 20, 30 agents in parallel, even me, I have a MacBook Pro M3 Max with 128 GB of RAM. Even with a $7,000 MacBook Pro, you still cannot run 50 plus agents on the same machine. So, the solution is

31:15 moving them to the cloud and not not just random architecture. And the solution is giving each agent his own isolated cloud sandbox. Right? So the first thing I tried is running

31:26 everything on a single VM, a single VPS where it's already better than running it locally. But you know once you run five agents or or 10 agents, you exhaust that VPS or that VM. So purely by uh

31:39 reasoning from first principles, it's clear that just like when you hire a new person at your company, you want to give each agent um isolated sandbox, its own computer, right? So that it can work, so

31:52 that it can do whatever it wants. So it has its own resources and it doesn't interrupt other agents. And this also allows for infinite scalability because it doesn't matter if you have 10, 100,

32:01 or thousand. It literally doesn't matter. Each agent has its own isolated sandbox. Not to mention, it's much better for cyber security because if one sandbox is compromised, you really don't

32:12 want that to be spreading to other agents and you really really don't want that to be on your main MacBook. So with the rise of hacks and exploits and breaches we've saw seen in the last, you

32:22 know, 7 to 14 days, it's becoming more apparent that whatever work you're doing with agents, it should be done primarily cloud first and then some agents locally. My prediction is that um it'll

32:35 probably be like ideal for 90% of the tasks. It will be ideal to run them in the cloud and maybe 10% local. You know, something that's like private or something that like, you know, you

32:45 wanted to analyze your your files in your downloads folder and organize some folders. Maybe it's better to run that agent locally, right? So you don't have to transfer a large volume of files to

32:53 the cloud. But I'm saying for 90 90% of work, it's going to be better to run it in the cloud. And again, I'm going to share with you my real data, which by the way, full transparency, even after I

33:04 started building Cloud Room, for a couple of weeks, it was still only around like 10 to 20% uh of agents I was running in the cloud. Why? Well, first of all, for me, it's always going to be

33:15 a bit lower because, you know, when there's issue with cloud, I need to fix it locally. So, just maintaining the software means I need to run more local worksheets than I would like to. But you

33:25 can see that in the last three or four weeks really uh the acceleration started and we've surpassed like 50% and even surpassing 60% and the trend line is clear like I can literally draw it um

33:39 where it's this is headed over the coming weeks and months and my prediction is that it's going to be like 90 to 95% cloud and 10% or like 5 to 10% local uh for me personally and for most

33:50 people as well because the reason and by the way as I started speaking to some people who were like, "No, I I really only develop locally." And as I started to unpack why, the answer was very

34:01 clear. People use agents locally when they experienced sub-optimal solutions. Because if you use some other cloud agent solution and they are missing something, maybe they're missing some of

34:11 your skills. Maybe the agents are missing an API key. The environment is not set up properly. It doesn't have your GitHub repos. It's missing some dependencies, some packages. Then you'll

34:20 be like, "Okay, cloud agents are overhyped. I'm going to use local. But it's not because your work is inherently local first. It's because whatever you've tried in terms of previous

34:32 cloudation solutions didn't work. It was a sub-optimal solution. Now, how do we compensate for this in cloud room, right? How do we ensure this is as good as possible? And why are we growing so

34:42 fast? Well, even when you run something as a cloud agent, so I can just show you. This is going to be a cloud agent. And let's say I forgot to give it something. can go to settings and

34:51 configure you know which skills should be preloaded which MCPS and API keys should be pre-loaded into every cloud sandbox even when I run it maybe I forget something right now if you give

35:02 the agent a task that it needs to complete and you don't give it all the files or or anything or something you block it and the agent will fail to complete that task purely because you

35:12 didn't give it a file you mentioned we solved it with this feature where you have full control over deciding how much access the cloud agents get. So you can put it to full like I have it. You can

35:24 turn it off. You can put it read only or ask. But these four levels let you control if an cloud agent is missing something that's only on your MacBook and you give him a task. Check the PDF

35:35 in my downloads and calculate how much I own on this invoice. If you give him such a task and you don't attach the PDF, the the agent normally would not complete it. If you're using Codex, if

35:45 you're using cloud desktop, if you're using cursor, this task would fail in all of these. Why? Because you simply forgot to attach the PDF in cloud room, this would not fail. I can literally say

35:56 how many files are on my downloads folder in my MacBook and what's the total file size? Check it and answer in short. So, even though this is a cloud sandbox and I can actually follow up

36:07 with a message. And by the way, where are we running? Are you running on my MacBook right now or is it in a cloud sandbox? Actually, I'm going to delete it so I'm

36:15 not leading a question. Is it somewhere else? But that was very fast. Right. Another thing with cloud agents, a lot of people don't adopt cloud agents because they're slow. If you used some

36:27 other product, I'm not going to name names, usually it takes 10, 20, 30 seconds for the cloud VM to boot up. We've done a lot of tricks and optimizations to make sure

36:37 that it's as fast as local. You just saw it for yourself. The agent started reasoning within like 1 second which is as fast as local. Um so not only does it feel as like you need to understand it

36:50 needs to feel as fast and it needs to have all the things right. So this I I asked it where are we running and again this is opus 5.5 the frontier model from entropic and says I'm not running on a

37:00 MacBook. I'm in a cloud uh cloud room cloud sandbox a Linux machine in the cloud and still it was able to access my downloads folder calculate how many files I had and say it adds up to 2.5 GB

37:11 purely because we figured it out with this feature and again we're the only ones in the world that have this feature of direct cloud to Mac access so that when you're running hundreds of cloud

37:21 agents like I am which in fact we can ask again how many of these are local and how many of these are uh cloud threats I would predict like it's probably half

37:33 and half roughly. We'll see. We're going to have it from the source so I don't have to guess it. Okay. So, it finished 553 open threads. 353. Okay. So, what's the percentage like 6535? What's the

37:50 percentage split? So, you can see that uh most Yeah. 64 36. That was a good mental math. 64% of my open threads are in the cloud right now and again you saw on the chart

38:03 is only increasing and I suspect it's going to increase when it approaches 100%. Because the reason when you have this direct Mac to access direct cloud to Mac access connection it really uh

38:16 the reasons start disappearing for running everything locally and it starts being more obvious for that you should run everything in the cloud in isolated sandboxes where each agent has its own

38:28 sandbox and they're not conflicting and they're not draining your resources and u it they keep running even when your MacBook is closed. This is like the most obvious thing and I should have

38:37 mentioned it earlier. The benefit of cloud agents is obviously they are running always right at 3:00 a.m. at at electricity blackout when you're traveling, you know, through a train or

38:47 driving somewhere where there's bad connectivity. If you have cloud agents properly set up with your GitHub repos, they can work all day, all night, no problem, with no

38:57 interruptions. Um, they always have internet, they always have electricity, they're always running. And that is important to building a business on top of agents. So

39:07 I think I touched on the most things. Again local it just means it runs on your MacBook. Cloud it means it runs in the cloud either on some VPS or ideally isolate sandboxes which is obviously the

39:16 right architecture which you know is what we have inside of Cloud Room. And again it's fully open source right I'm not like trying to gateep it. I literally gave all these inventions away

39:28 and uh we're the first company to figure out how to run cloud agents with any harness, any model and any interest provider fully open source. So again, cloudroom.dev link below.

39:40 Let's talk about the AD agent development environment. There's a lot of them. How to pick one, right? Well, think about it again. What matters? What are the characters correct

39:50 characteristics that matter? Well, you need every top harness, every frontier harness, whether it's cloth code, codex, pi, hermes, droid, devin, open code, OM, whatever harness you prefer, you need

40:03 them to be available. If you're using maybe, I don't know, a lot of people are using cloud desktop. Sure, it's good for using anthropic models, but you don't have anything else. The moment you want

40:12 to try the latest Deep Seek, the moment you want to benchmark the new news wonder, the moment you want to do a GP review, you cannot do it in cloud code, right? You cannot do it in cloud

40:23 desktop. Same with Codex. Codex was great at the start of the year. The Codex app was one of the best in the world in terms of the UI. You cannot use Fable. You cannot use

40:32 Opus. You cannot use any other model except for OpenAs models. This is really the reason why your AD cannot be from the labs. And by the way, when people are saying, "Oh, David, why aren't you

40:42 scared of OpenI or Entropic? Couldn't they just replace Cloud Room?" Well, guess what? Just the fact that first of all, they're not open sourcing their stuff. And I'm fully open sourcing it

40:52 and a lot of people are, you know, liking that because you can see the code, you can do your own fork. We have over 230 forks and we just launched like two weeks ago. It's crazy. But the main

41:03 thing is they will never open it up to other models. They will never let you use other harnesses and other subscriptions. Enforopic will always going to protect itself and only let you

41:12 use cloud models. OpenAI in Codex is never going to let you use any other model except for GPD models. It's very simple, right? Like you don't have to be a genius to realize this. Next thing is

41:24 your own subscriptions, right? The issue with some of these like third party whether it's like Cognition, whether it's Conductor, Windsurf, um I guess not Conductor, but like AM

41:34 and Cursor, they force you to use their own subscriptions and their own harness. Like it's ridiculous. If you're paying $200 a month for CL Max or GBD Pro, you want to use that subscription. So,

41:47 whatever AD you use, make sure it allows you to use your tokens you're already paying for. Another thing is agents run agents. So that's another big limitation is u I can just show you a quick demo.

41:59 Launch f sub agents to um do more analysis of all of these threats and let me know how they split split up across categories of work. Don't reveal too much. I'm recording a YouTube video.

42:08 It's going to get tens of thousands of views. So just tell me how it u how it's split out and don't do the analysis yourself. chunk this um 553 threats into 20 separate sub aents each one running

42:20 haiku 5.5 extra high inside of clo code you cannot have agents starting another top level threats you can maybe have it start sub agents but if I want to say like in the

42:33 same folder we're in right now um start five other top level threats but each of them should be codex and it should be GPT6 Astra and just tell them to answer what machine they're on to not

42:45 make any changes. Again, same project. Make sure they are top level cloud threats. Okay. So, here this one is launching the sub agents. So, let me show you how easy it is. And these are

42:53 local sub aents. I'm going to have this agent launch a tople cloud threats. So, you see both local and cloud because you know people get confused when you see like cloud room. I want to still do

43:03 local work. Well, guess what? It's still the best AD in the world for local stuff as well. So, this is a big limitation. for example in cloth desktop where for some reason Enforopic doesn't let you

43:15 tell threats CLI control. So again here it is this is a Opus 5.5 and it's launching Astra top level cloud threats. Okay, this is not possible in most ads because they don't have good CLI

43:29 controls and they're not really agent first and models can talk to each other. You know, they can use room CLI. That's the CLI we have room CLI. I can actually

43:39 have this explain room CLI. So, it's a skill that allows any agent to control the CLI, allows them to check any threat, allows them to start new threats, archive other

43:51 threats. So, right, for example, here it's running the 20 sub agents. And I can say now stop all sub all of the sub aents. Boom. So this opens 55 main agent and

44:02 again these are ha coup just to check just to prove it to you. Just like that I can tell him to stop and he can literally interrupt all of them and they will stop. Boom. They're no no longer

44:10 running. I can say now archive all of them and I can presend it. So it like interrupts it and steers it. Obviously we have all the classic features like steering like cute messages all that

44:20 stuff. But the main thing I want to stress is that when you're picking an ADE, it needs to be fully controllable through CLI so that agents can pro programmatically control it. They can do

44:31 any other operations that you yourself can do because in the future, who wants to click around and like start 20 different threats like what is this manual labor? This is the modern day

44:40 working in a coal mine. You don't want to just tell an agent start 20 sub agents or start five top level threats. right here this guy Opus 5.5 running in a cloud sandbox isolated cloud sandbox

44:51 launched five different cloud tople threats so these are not his sub agents they are also top level and they're using a different harness codex harness with Astra as the model right complete

45:02 configurability everything could be agent first everything is available through CLI everything is fully compatible both local and cloud be careful what AD you choose obviously

45:14 um I think cloud room is the first one and our growth backs it up with how fast we're growing. But try them yourself. Again, as I said multiple times, Cloud Room is free and open source. Just

45:23 download it, try it, give it a shot, use it for a few days, see how much better it is than cloud desktop and Codex. Um, I think you're not going to go back. Now, let's go to the next level because

45:35 to unlock this productivity as I showed you with this with the chart of growing commits and growing productivity, you need to make sure you remove yourself. What does that look like? Agent loops.

45:49 You need to start identifying highly repetitive and predictable parts of your own workflows and how you work and turning these into software factories. But software factories is too

45:59 intimidating of a firm of a term. A lot of people hear a software factory and they like their heads start spinning. They don't know what what's happening or how to create a software factory. The

46:08 way you create one is you actually don't want to create a software factory. You just start by creating small predictable agent loops for the most boring and repetitive work that you do. That's how

46:19 you start. And once you have 10, 20, 30 agent loops running at all times, there you go. Suddenly you have a software factory. But you don't get there by trying to build a software factory. You

46:30 get there by creating one or two agent loops at a time. So the big productivity increase especially over the last two to three weeks for me like really here going from like I don't know what is

46:41 this 400 500 contributions per week to like 700 nearly thousand for this week is agent loops because you you you only have so many hours in the day. Trust me guys I'm working literally from 9:00

46:54 a.m. to like 10 11 p.m. every single day 7 days a week. There's only so many hours in the day. So if you want to be on the cutting edge of AI and really be fully squeeze your potential to the

47:06 fullest as an individual, you need to be automating parts of your work. And the way to do that is with agent loops. So let me go into that. A lot of people have no idea how to

47:17 create agent loops or what they are and when like when is a good place to create an angel loop. So first of all, you should completely stop just blindly uh copying other people's work.

47:27 and other people's skills and uh loops. You should have clear understanding of when to copy over a skill, right? So don't go and like take all my skills blindly. No, you should go and find two,

47:39 three, four skills that are relevant to you and implement those. Same thing with my agent loops. Start with those that you really need. Don't just go and copy. Like there's so many people who just go

47:49 and copy everything Matt PCO does. Why are you Matt PCO? Is your day-to-day exactly the same as his? Because unless you're literally his biological clone, his twin, why are you copying all of his

48:02 skills? And again, the same thing goes for my skills. The same thing goes for any other person's skills or agent loops. You should be doing them and creating them and using them to solve

48:11 specific tasks and problems in your own life. The skills that me and you are going to use every single day are going to differ slightly. Maybe some of them are similar if you do software

48:21 development like I do. But it should not be onetoone. And the same thing goes with agent loops. So let me explain um with a simple Minecraft analogy what tasks you should create

48:33 agent loops for and what you should not create agent loops for. So to really explain the type of tasks you should try to automate fully end to end with agents. I'm not saying like just human

48:44 in the loop. I'm saying no human in the loop fully automated. It is kind of like the farming wood analogy. Okay. So, anytime you want to do a big project, and I'm using Minecraft because it's the

48:54 most popular game of all time, you need some materials and resources, right? And chopping wood, there's nothing there. Like, you can just chop wood. There's no prestige or pride in shop chopping wood,

49:04 doing it manually. So, this is like a boring, repetitive task. It's very obvious. It's always clear what to do. Building a cathedral, on the other hand, is extremely

49:15 difficult. Most players, most people will never achieve anything close to this and it's difficult. It takes a lot of creativity, a lot of planning, a lot of

49:25 high level thinking. If in this game or any game you automated this part, people would call you a cheater, right? If you automated this this meme of like, oh, just let me do a little mining of camera

49:35 and then comes with a chest of diamonds, right? Doesn't matter if it's diamonds, if it's wood. If you automated this in a video game, you would be called a cheater. But the beauty of business is

49:45 as long as that is legal, you can do anything. If you automate half of your work, if you automate 80% of your work and your business is running and growing faster than ever, you're it might feel

49:55 like cheating. And that's what it happen. That's what it feels like when you really unlock AI coding and when you really embrace agentic engineering to the fullest and you figure it out. You

50:04 can focus on building the cathedral, right? On these tasks that are highly creative, that are like still uniquely human. It takes a lot of planning, a lot of practical thinking, a lot of abstract

50:15 innovation and invention and taste and you can fully automate chopping of wood, right? If there's a bug report and it's a very easily reproducible bug that Opus can identify, why not automate this end

50:29 to end, why not put an agent loop with Opus there and have it fix it, right? Inside of Cloud Room, we have loops built in, so it's as easy as possible. Right now I have 33 different loops

50:40 running. That results in about 2,000 runs per day. 86% of them run on their own, which means when there's a bug, Opus runs it. What what happens after that? If it identifies, okay, it's a

50:53 real bug. It's a serious thing. Reviews it with GPT 6.1 soul telling it like adversarial review. Is this a real thing? Is this actually a bug or is this overthinking? If the

51:06 other frontier model from a different AI lab also says, "No, this is a real bug. It's affecting users. I was also able to replicate it and yes, the fix is clear and safe." If you have two frontier

51:17 models from different AI labs, both agreeing on this, what are you doing? Like there's nothing to be involved in. This should be fixed and shipped straight to production. This is your

51:28 equivalent of chopping the wood. You need to start identifying these types of tasks, right? Again, security sweep there. If there's a if there's a feature suggestion that like 70% of your paying

51:41 customers suggested, maybe you should have Opus build it and like present you with three different variations and you se select which one to do, right? And again, you can have loops that are fully

51:51 automated, fully autonomous. You can have some loops where you are still in the loop. It can be conditional like these agents are smart. Again, frontier models, guys, please. You see, I'm using

52:01 Opus. If I want to be a little more cost efficient, I'm using Sonet. But most of my uh agent loops are running Opus 5.5 or uh GBD 6.0 soul, right? This one is alternating. That's another beautiful

52:13 thing. If you have agent loops alternating, these models find different stuff. And I told you earlier, I'm going to explain why I use different frontier models. Even though like one of them is

52:23 the best one right now, I think Opus is the best one. The reason for that is like asking five of your smartest friends for feedback. Even if one of them is the smartest by far, if you ask

52:35 five, each of them is going to see something else. They're going to spot a different mistake that even the smartest guy missed. It's the same logic with Frontier models. If you ask Opus and

52:45 then Soul and then Astra and then Muse Wonder and then Gemini Argan and Kim 3.1 Pro, each of them is going to find something else. And if if not, sure, you're lucky there was no bug or no

52:58 issue. But let's say there is, right? Like you really wish that you ran the other frontier model. So back to this analogy, you want to identify what are the boring

53:09 things like web security test. Do you really want to check your own web security every four hours? No human developer is doing that, right? All of these larers, they're like, "Nah, what

53:19 I'm doing AI could not be used. I'm really writing handcrafted enterprise code. Guess what guys? All of these security issues that we're finding come from human written code. These PHP

53:31 issues and these like zero days, they've been there for 12 years. People just haven't found them. And now with running swarms of agents, they found them, right? So the most important loops um

53:42 are obviously like cyber security and all that stuff. But the reason why people don't adopt loops and what I'm trying to explain here is because it's a big big leap of activating. There's a

53:52 lot of friction before you activate your first agent loop, right? For me, when I have like 30 plus, it's easy to add a 31st, 32nd, 40th, 50th loop. The hardest part is going from zero to one to two

54:04 loops, right? Because you might be like, "Okay, David, it's cool you're running all these loops, but I don't know. I don't know if I need them for me, you know? I don't know what they're doing. I

54:12 don't know how to implement them." Well, guess what? You need to identify it for your own work. Your first one or two loops should be from your own stuff. And how how do you do that? Well, you

54:23 analyze your own work because if you're using cloth code, codeex, Hermes, PI, open code, you have all these sessions. And that is a gold mine because those sessions

54:35 have embedded your usage patterns. Those sessions have your prompting habits, your sequences of of skills you run, how you work, how you solve problems. And by the way, uh you can do this for very

54:48 cheap. I invented these scans which all of you can use inside of Cloud Room. Again, top left, go into loops and click on scans. You can analyze this for less than 1% of your weekly limit. So uh you

55:00 know whatever you use cloth or codex these scans are very cheap because again it chunks the works into parallel uh haiku sub aents because this is the perfect type of task for sub aents if

55:13 you were paying attention earlier it's predictable and it's easily chunkable so whatever this looks like for you maybe you don't you have more closed code and codex maybe you have more open code or

55:22 pi or her whatever doesn't matter whatever your sessions are on your computer you can analyze them with one of scans. So, I'm going to show you what that looks like. You can just click on

55:31 scan my workflow. And this launches an Opus 5.5 main agent. Where did it go? I think it went into pint. There it is. And he's going to uh start uh 20 to 30 Haiku sub agents, right? Haiku is super

55:44 cheap model. That's why this only uses less than 1% of your weekly limit. And it's going to analyze thousands of sessions, which for me, again, I showed you was like 3,000 sessions on my local

55:55 and obviously I have much more on my cloud. and it's going to look for the first one or two loops to suggest specifically for you. So, right, not using some other, you know, person

56:07 online or entrepreneur or influencer. No, this will analyze your sessions and it's going to look what habits and workflows you're repeating. And not only that, it's going to identify the 9010

56:19 split. So, I'm going to explain that, right? Let's say from this comes out two very clear things processes you're repeating right and these are the two suggestions for a loop instead of just

56:31 activating these loops and trying to automate that fully I architect architected it so it's in a 9010 split what does that mean let's say 90% of this is still like kind of risky changes

56:45 or requires a lot of taste lot of u judgment a lot of context that maybe you don't trust agents when it analyzes thousands of your sessions or maybe hundreds if you're

56:56 newer to AI, it will look for the 10% in those two. So again, it it is instructed the scan the way I built it. It's instructed to find two very specific things, the very specific like habits or

57:09 very specific repeatable workflows that you are doing and then suggest how this could be turned into your first two agent loops. And again guys, I'm trying to make this as simple for you

57:19 because I see where the future is headed and I also see where most people are at and most people are not running any agent loops. That's why inside of Cloud Room, which again I have to repeat it,

57:30 it's free and open source. You can literally go to cloudroom.dev. It's going to be the first link below and just download it. And yes, we do support Windows now. Download it, get it up and

57:39 running, use whatever harness, whatever is your favorite model, doesn't matter. I'm trying to make this as easy for you as possible to see the future and to claim the benefits of the future. So,

57:49 I'm I've added the loops in the top left. And not only that, I now introduce the scans where it's couldn't be easier to identify what your first two agent loops should be. And you can see that

58:01 this is very fast. It's very efficient because again, Opus is the main agent and it doesn't actually um run everything himself. He starts the Haiku agents and chunks the work into 31

58:11 different slices. And these 31 sub aents analyze u those sessions to look for your patterns and your your repeatable workflows. But the real innovation here, why this is useful is that it doesn't

58:22 just tell you, yo, John, these are your to top two things you're doing. Automate them fully. That is a big ask. No, it does a 9010 split. It identifies, actually, I call it 1090 split. It

58:35 identifies the 10% that is the safest, the least risky, and This is where the loop runs autonomously. For the other 90%, the loop analyzes it, researches it, and and

58:48 presents you what it thinks it should do, but waits for your opinion. That way, you can enable your first two agent loops. Get them running without fearing like, oh, what if this agent messes up?

58:59 What if it doesn't do it in the way I want? What what if this loop suggests some slop? I don't want slop going to my production. Nobody does. Okay. The reason I can ship so fast and why I have

59:10 a public change log where you can see changes literally coming twice a day, which by the way, this is a pro tip. Anytime you see people flexing how many PRs or commits or markdown files, god

59:20 forbid. When I heard Gary Tan mentioning that he has half a million markdown files, I I cringed. I This is incredible. People flexing how many markdown files they have. It's It's

59:29 beyond Like, what is that? What is that measuring? How much slop you're creating? You need to ask people, where is your public change log? Show me what you're doing. I have a public change

59:38 log. You can see exactly what I'm doing every single day. Fuller transparency so you can see the changes. And I would put my hand in the fire saying that each version of Cloud Room is better than the

59:50 last. So next time somebody's flexing on Twitter that they're running 10 billion sub agents and are shipping thousands of PRs per day, ask them where is their change lock and why it isn't public.

1:00:02 That's a very important test. So the way it works, I don't know if it's finished. So, okay, for me it says most of your repeat work already runs as loops. Feedback, triage, side effects, reviews,

1:00:11 PR reviews, publish, blah blah. Yeah, I already have over 30 loops running. So, it's [clears throat] going to be a bit bit tougher for me to find it. But this, even for me, when I ran it last time, it

1:00:21 suggested two brilliant new loops. Let me see. Yeah, here these two. I'm I need to show you this. So, this is what it suggested for me last time. Loops I would turn on. And this is what it looks

1:00:31 like, by the way. The 1090 split 10% when it runs on its own. And guys, the first time it runs, the reason I'm so excited about this, it's completely amazing feeling, fully automating some

1:00:43 part of your daily life, you literally reclaim valuable minutes from your life back to you. And it's specialized and personalized to you. So for me, even after I had over 25 loops, the moment I

1:00:56 ran this, it found two new opportunities which I did not have. The first one being release check daily at 9 uh a.m. Pacific time and old Fred sweep daily at 7:00 p.m. Pacific time.

1:01:10 The first one checking for a new release okay of a new release for a version because I do this once or twice a day as you've seen in the change log. And then the

1:01:20 second loop checking for old threats going through them and seeing if some of them should be archived. And the beautiful part which I you know came up with is the 10 and 90 split where it

1:01:31 looks for the 10% of the most safe and obvious chunk of that loop that can run safely fully autonomously and the other 90% waits for you. It waits for your input for your confirmation when it's

1:01:43 not sure it doesn't ship any slope. And this is how you guys can activate your first two agent loops today and have fully autonomous work being done. I don't care if you're in marketing. I

1:01:54 don't care if you're a software development. I don't care if you're in sales, if you're in research. Doesn't matter. Whatever you're doing, you have some repeatable parts of your workflow.

1:02:05 Chances are you probably just didn't analyze and scan your agent sessions, which again I turned this into a one-click feature trying to make it as accessible as possible. Um, all of this

1:02:16 is available inside of Cloud Room. And this scan runs. Let me see the results here. Okay, so it suggested the same one as last time and a new one. Morning focus. Okay, interesting. So this is the

1:02:28 same release check and then it suggested morning focus. This is something I started doing the last three days unpinning some of my pin threads because I have a habit of having too many pinned

1:02:36 threads inside of Cloud Room. So yes, this is actually stuff I'm doing. It doesn't hallucinate. It doesn't tell you, yo, you need to enable these five random loops or like this influencer has

1:02:47 released his stack, you know, stack of skills. Download it. Why? These are not your skills. You don't have the same habits as as that random influencer. You need to make sure that whatever you're

1:02:57 doing, whether it's agent loops, whether it's skills, it is highly relevant to you. So do it. Scan your sessions, analyze them, see what patterns and what the behavior is repeating. And then

1:03:08 based off of that, enable the first a two engine loops in the 1090 way where 10% of the most obvious and safe situations it executes fully autonomously. and in the other 90%

1:03:21 it waits for you. Trust me, when you do this, it's absolutely liberating and it's amazing and you really start reclaiming hours of your day. So whether you're an entrepreneur or not, time is

1:03:30 money. And if agents can do even a small percentage of your work fully autonomously, you can use that time to do more productive stuff or, you know, go out on a walk or or to the gym and

1:03:41 improve your life in some other way. It's entirely up to you. So now that I showed you how easy it is to create your first one or two agent loops, let me explain the the philosophy and the

1:03:53 ideology of loops and also how people like Dex um do it and how not to do it and how to how to do it for more than software because a lot of people still have this limiting belief that you can

1:04:03 only run agent loops for software engineering. That's completely not the case and I'm about to show you how you can run it for all parts of your business and life. Some of this I

1:04:11 already explained to you, right? Uh, I have around 30 loops, cyber security, obvious fixes, bugs, customer customers find. So, I'm not going to repeat that. I just want to make sure you understand

1:04:22 you get rid of all your limiting beliefs. If you're not fully embracing AI to the fullest, chances are you have multiple limiting beliefs. Either you think AI is not that good or you think

1:04:31 your tasks are some like special unique tasks that even not even 10% of them, not even 20% of them AI could do. or you think like it's all over and AI is just replacing

1:04:43 everything. All of these are completely wrong. Okay. The best way to understand the future is to learn from the past. Every revolution people stopped doing the task and started working on the

1:04:53 machine. Right? So if you we look at the 1830s textile mills instead of weaving cloth by hand you were running the power looms.

1:05:03 This was massive productivity increase. allowed people to create cloth and way more clothing way way faster, way more efficiently and also way cheaper. 1913 Henry Ford, shout Henry Ford, the

1:05:14 assembly line, right? Instead of building each car by hand, you were running the line and people were specialized to build just the wheel or just the chassis or whatever. Um, and

1:05:24 the role of the entrepreneur or the car builder was to assemble the machine, the assembly line. 1946, human computers were literally floors of men and women, like 20 floors in a

1:05:37 skyscraper doing math by hand, pen and paper, slow addition, multiplication, division, subtraction, extremely slow and inefficient. Then that changed into using calculators and and programming

1:05:49 the machine. And guys, this this is like we always think that we are so special and that our time is different. It's not that different. Okay, will turn these specific things that right now if you if

1:06:01 you take the red pill and you go beyond and you're using AI for everything, it really is clear what is manual labor. Filling out some form online, doing research manually, like what are

1:06:12 you talking about? Like clicking through websites or doing research manually instead of using something like deep API to do a deep research absolutely unacceptable.

1:06:20 Guys, are you checking your the entire codebase deeply for security issues every 3 hours? No developer is doing that. I have daily report on a funnel. I have tweet scout for people mentioning

1:06:33 my product. I doing customer outreach um reminding me like what to which people I haven't contacted. Bug reports every minute. All of these things have either fully automatable workflows or parts

1:06:47 like big parts of these you know tasks that can be fully done by agents. Even when you have something that like you still want to decide whether a feature is shipped maybe the first 80% is done

1:06:59 by AI designing the feature implementing the code reviewing it with another frontier model showing you three different UI designs. You just come in, you pick one of the UI

1:07:08 designs, you ask one or two clarifying questions and you you approve it or you come in and you don't approve it, right? [clears throat] It's completely fine. But the point is just like every

1:07:18 previous industrial revolution, we will stop doing specific tasks when it comes to software and knowledge work and we will instead be running our machine. We will be running the software factory.

1:07:28 And again, you don't start by uh trying to build a software factory. In fact, if you watch my video with Dex, you know that it's a complete disaster. He tried building a software factory for for four

1:07:39 months that nobody read and it was just slop and angry users. It was terrible. You never start by trying to make a software factory. You start by creating your first agent loop, your second agent

1:07:49 loop, and sooner or later you have 20, 30, 40, 50 agent loops running and there you go. That is a software factory. My point with this point is that um it doesn't need to be software. It can be

1:07:59 marketing. It can be uh outreach. It can be research, it can be uh GTM, whatever. Anything, any type of work has some parts are that are repeatable and these are clear opportunities to build an

1:08:12 agent loop inside of them. Now, this is what I promised you earlier showing you how I built self-improving software. Cloudroom is self-improving, right? Users, which a lot of you guys are the

1:08:23 users, when you report a bug, every single bug gets analyzed by Opus, right? within like 10 to 30 minutes to make sure that it's, you know, a legit bug, that it's not just some spam. A lot of

1:08:34 people use agents to spam open source repos. I built a lot of protections against that. But the main point is you want to make sure that like if you have real users, you want to respect them.

1:08:42 You want to respect their ideas and and if they run into errors, you want to fix those errors as soon as possible. And you can do a lot of that by building self-improving software. So, what I've

1:08:52 done and uh what I think like I haven't turned this into a feature yet, but if a lot of you want it, I might turn it uh into a feature. I'm going to have a agent explain it. Use room CLI and

1:09:05 explain how the feedback feature works. So, what we have is actually you can click it from the button right here. Report a bug. Use the feedback tool. Report the following items to David.

1:09:17 Right? So, not only you can do it, but your agents can do it, right? So if you encounter a bug or your agent encounters an error, he can report it directly to me. Now again, I haven't turned this

1:09:28 into a feature so that others can people can use it, but if there's demand, we can make it happen. But the core idea is that I want to reduce the time between somebody encountering some bug or error

1:09:38 or it doesn't have to be a bug or error. It can be a suggestion. It can be an idea. I want to reduce the friction and the time of you as a user getting it to me. So whoever is using my software

1:09:49 which again could be cloud room could be deep API or something else I want you to have the best experience possible and if you encounter a bug it should be less than 24 hours before that bug is fully

1:09:59 finished and the way the only way to do that is by building a self-improving piece of software. So again I try to extract it the way it works into a clear diagram and uh basically what happens is

1:10:10 that anytime you report something through the room CLI feedback and you don't have to know any of these CLI commands that's the beautiful part you simply tell your agent in plain English

1:10:19 right it's like yo today cloud room has been feeling a bit slow report this to David or maybe this type of attachment didn't work make sure to tell David and the developers you just tell your agent

1:10:31 in plain English and he can report something and it automatically opens a thread for me in my own cloud room. That's why I have so many threads because uh I'm actually analyzing what

1:10:42 my customers report to me and what bugs they encounter, right? And uh an agent automatically starts on it. And again, if it's something simple, obvious reproducible, it can be shipped directly

1:10:55 straight to production, maybe with a check with another Frontier model. If it's like a feature suggestion, a major product decision, the agent starts on it, explains it to me in the format that

1:11:05 I want and I make the decision. But again, it's kickstarted and it got moved into where I work, right? From an agent report, from a human, a user reporting to his own agent. the agent using a CLI

1:11:17 tool for it to go to my Postgress database table and that gets triggered from an agent loop into my own workspace where I work which you know obviously I work with agents inside of cloud room

1:11:30 that cuts so many steps and so much time and allows me to build self-improving software that ensures that every single um you can see here like most of the most of the changes it's not just like

1:11:42 new features a lot of these are like improvements and and bug fixes and edge cases that our users found because every single version is getting better because I've built this self-improving software.

1:11:52 So, let me break down how it actually looks like, right? How Cloud Room improves itself. And you can take this, feel free to copy this and implement this. Um, I mean, everything from this

1:12:00 video, right? Feel free to just take it and copy it for your own purposes. So, here's how Cloud Room improves itself every single day. every single report or bug or you know suggestion turns into a

1:12:09 threat um on my own cost on my own inference and then the ones I agree with I approve it and it ships or one of my developers obviously right um if you have a team you can divide the the work

1:12:21 accordingly so anyone can report it and we have a we have a checks to make sure that you know different people get their own reputation if somebody is a is a scammer and a spammer they get low

1:12:34 reputation and Okay, you know, the agent is more strict with seeing like, yo, this person, they've been spamming. Um, they're just trying to get contributions.

1:12:44 You know, probably their changes aren't in good faith. But if somebody is like only reports good bugs and has like good intentions, people from the way people like create issues or pull requests, you

1:12:53 can very easily tell if this person just wants to be on the contributors list and is trying to find random or if this is a, you know, real user that found a bug in his real workflow. It's not that

1:13:05 hard to tell and in fact it's very hard to fool the frontier models like Opus 5.5 with extra high or max reasoning effort. So anybody reports user reports it either through the button or their

1:13:15 agent or the agent hits a bug and reports it by itself depends then it goes to a loop that runs every minute if there's a new report right again check through my posgress database table if

1:13:26 there is it starts as a new cloud thread and the beauty is this guys it's all explained in this diagram it's really beautiful so yes it creates a new thread in my own cloud room agent investigates

1:13:36 it read only no edits now if it's a real thing a real product improvement a real bug, something serious, it goes as into a GitHub issue and gives me a short summary. If it's not a real

1:13:49 thing, either noise or something duplicate that we're already handling, the thread archives itself, it's not bloating my cloud room because if it's not a real thing that should be fixed,

1:14:01 opus self archives. We literally have this skill run selfarchchive and any threat can archive itself. Boom, it's archived and it's no longer in the left sidebar. So

1:14:12 this makes sure that everything gets investigated and if it's a false alarm, maybe the user is using an old version and it's already fixed in the next version. I'm not wasting any time on it.

1:14:21 Any human developer should not waste time on it. If that bug has already been fixed and you know it's just coming in the next version, it's automatically archived. It's really beautiful. Like

1:14:30 when you have these systems, it's just amazing. So when is a real thing, real bug or a good idea that hasn't been fixed yet, it goes and waits for me. Then I do I approve it or not? If not,

1:14:43 it's archived. If I do approve it, the agent fixes it. We review with a different frontier model and ships it. Right? Better cloud room for every single user every day. And uh the next

1:14:52 report goes back. And this is really the self-improving logic of how Cloud Room is becoming better every single day. And again, I'm very transparent with what updates we're shipping and what we're

1:15:05 releasing. And anybody who uses Cloud Room knows just how quickly it's uh it's improving. And anytime you encounter any serious issue, it's going to be like usually within 24 hours, but at the most

1:15:18 48 hours until it's fixed because of these systems, right? So, and feel free to uh rebuild this again. This is full transparency. I haven't turned this into like a feature where you can like plug

1:15:29 and play self-building just like you know I made Asian loops really easy feature to create a new Asian loop or the scan. I made it a super one-click feature that you can just implement for

1:15:38 yourself. This I haven't turned into a clear product yet. If there's interest for this self-improving software, u I'll of course make it. Um, so that's that. Now, let's talk about self-hosted

1:15:51 inference local models, when to run them, how I'm running them. Um, yeah, a lot of people are interested in that. All right, so self-hosted inference, local models, how to use it, when to use

1:16:01 it. You saw that I'm using Frontier models mostly, right? Well, I still have a local setup. Right now, I have two DJX Sparks and I just bought two more. So, I'm going to have four DGX Sparks for a

1:16:14 grand total of [sighs and gasps] half a terabyte of RAM, 512 GB of RAM. I also have 128 on my MacBook Pro M3 Max. I also have 20 GB of RAM across my two iPhones and another 24 GB in my uh

1:16:29 MacBook Air. Now obviously I'm not chaining all of these but the reason I'm saying this is I still think local models are valuable okay local models are still highly valuable just you most

1:16:40 people overindex like there's nobody with a nuance inter nuance opinion right literally you go on Twitter you go on YouTube and people are either local maxing and they're like oh you can run

1:16:51 Opus 5 at home blah blah blah or there's people who are like ah self is useless um everything will be in the cloud whatever so first of all if you're only If you don't have any um local compute,

1:17:03 you are simply a slave. What does what does it mean? Well, if somebody turned off your chbd subscription or your enthropic subscription, you could not use AI, right? You could not use any

1:17:15 models. Maybe you could run something small on your laptop. But unless you have 128 GB of RAM, even that it's going to be very small. You need to run you need to like own local inference. You

1:17:26 need to own your own hardware for the same reason. You need to have like solar panels on your house, right? Or you need to have your own garden.

1:17:36 Again, not every person can have a garden. Not every person can have solar panels. But I don't care. Like I'm not making this content for people who are like beginners. This is for those who

1:17:46 want to be on the cutting edge of AI. If you want to protect yourself in the future, the most valuable thing resource will be intelligence tokens. Every single month, you can run better

1:18:00 and better models on the same hardware. So, with these two DJX Sparks, what I could run 6 months ago was not that good. But now, I can run the latest DeepS V4.1 Flash 1017, and it's

1:18:14 incredible. With four DGX Sparks, I'll be able to run either more instances of the same agent or the same model at higher speeds or a bigger model. The thing you need to

1:18:25 realize is that the hardware is growing in cost. But what's growing even faster than the cost of the hardware is the value you can generate from each hardware. In fact, if you see the

1:18:34 graphs, I think A6Z posted some interesting graphs. The same GPUs both for consumer GPUs and data center GPUs have drastically grown in the value of their output tokens over the past six

1:18:47 years. So people thought like A100s which is like you know two generations old or whatever from Nvidia they would not be really used. Actually it's three generations old. They they thought they

1:18:58 would depreciate after three or four years. Guess what? Like six years later they're still growing in their output value. So you need to own hardware to simply protect yourself. If there was

1:19:08 some tyrannical government, if there was some issue with some of the labs being nationalized or whatever, you need to make sure that you still can run your own inference at home. Sure, maybe you

1:19:17 cannot run a model that's as big as Fable or Astra, but you need to run something to protect yourself. Something that you can run fully locally, fully privately. This is really uh the like,

1:19:29 you know, downside protection when it comes to being practical. You do this for maximum privacy. When you're reviewing like your health or some like legal contracts or whatever, this is

1:19:39 where you want to run a powerful model that's maybe 90% as good as the frontier, but it's fully local. It's fully offline and you don't send any of that private valuable data to enthropic

1:19:50 or openi, right? So the way we use these DGX sparks is anytime there's like a PDF with a legal contract anytime there's something that like has sensitive data you know employee data whatever we could

1:20:03 we send it to our deepseek which is hosted on these two sparks and um none of that data goes to the cloud none of that data goes to open anthropic Google um you know XAI muse whatever so does

1:20:19 that right you use it for private stuff Or you can run unrestricted models which is a whole separate video which don't you know which basically answer anything um which again for some use cases like

1:20:30 if you want to go to political topic or some like more controversial topics JGB and CL are going to give you a very very biased opinion you know like Grock is is a bit better but still it's heavily

1:20:42 biased. So sometimes you just need to ask a honest question and for that you want to run a obliterated or unrestricted model and you need to do that locally. I mean there's plenty of

1:20:51 APIs like Venice is a good one. So you can still do it in the cloud but local inference is more of a practical thing like if something happened in the future you want to have some hardware so that

1:21:03 you can still use AI and protect yourself. And by the way in all of these famous hacks where somebody was attacked by AI model the frontier models from open AI anopropic they start rejecting

1:21:14 they don't let you defend yourself and people have to go to like GLM or Deepseek you know these open source and open weights models to defend themselves. So uh self-hosted inference

1:21:23 right you don't use it for the most frontier work right for difficult coding and and difficult tasks obviously you should use the best model possible opus 5.5 GBD6 Astra GBD 6.1 soul Gemini 4

1:21:37 argan or carbon whatever is available okay Frontier model for most of your work anytime you're dealing with something highly important highly sensitive highly private that's where

1:21:46 you want to have ability to run a local model the problem is if you don't have uh powerful hardware if you don't have multiple sparks or the the new Mac Studios is crazy, right? Like the uh Mac

1:21:59 Studio M5 Ultra 512 GB with I think like 1.2 terabytes uh per second bandwidth. It's in in like incredible just insane. um same amount of RAM as four sparks but like six times

1:22:17 or five five or six times more bandwidth allowing for faster inference. So if you have only like your MacBook or only like some computer, it's very limiting in terms of how powerful model you can run

1:22:30 locally and then you're probably not going to run it locally because you know the the delta between the frontier models from open anthropic and whatever you can run is going to be so big that

1:22:39 you're going to be just pessimistic. But if you invest some thousands of dollars um or tens of thousands of dollars if if you're really serious about AI into hardware then uh even if there's a World

1:22:50 War II even if OpenAI shuts down or anthropy goes under or the US government bans export of tokens to outside US citizens no matter what happens you are protected and even if you have something

1:23:01 highly sensitive like the health of your child or something something you really don't want to send to openropic you can just discuss it fully at home local inference right so that is the practical

1:23:10 answer. Um, again, I'm I'm fully transparent. Most of my work is done with Frontier models. Um, we have this office inference. We're using it more and more. I I told you I

1:23:22 have two sparks right now. I just bought two more literally two days ago. So, they're coming next week. So, I'm going to have 4 DGX Sparks and I'm probably going to buy the new Mac Studio M5 Ultra

1:23:31 with 512 GB of RAM. The only problem is that it's going to be starting at like 22 23K USD. So, it's going to be very steep. But, by the way, if you think the hardware prices are bad, they're only

1:23:44 going to get worse, right? So, if you're waiting for a drop, you're going to be you're going to be waiting for God. You're going to be waiting for a long time. Um, the sparks I bought are

1:23:53 already like 30% more expensive than I bought them a few months ago. So, yeah, that's that. And, uh, use local when it makes sense. Use the Frontier model when it makes sense.

1:24:04 Don't be one of these absolutists that thinks like everything needs to be local or you should only use openly anthropic and local makes no sense. I don't know why most people are incapable of

1:24:14 critical thinking and having a nuanced opinion. Obviously, there's use cases that can be done fully local and obviously there's many tasks where you want the smartest model possible. Just

1:24:24 use whatever makes sense. And again, if you watched until the end, you're probably serious about AI. So, give Cloud Room a shot. Okay? As I mentioned many times, it's free. It's open source.

1:24:35 You can try it um on any platform, whether that's Windows, whether that's Mac OS and Linux. We didn't have Windows support. Now, we have it. Okay? So, if you're a Windows user, please download

1:24:44 it. Give it a shot. Let me know if there's some friction. If there's any issue you encounter, just say, report this to David. Okay? Or if you're like chatting, you know, here in any chat,

1:24:56 you have this button top right, report a bug. Boom. The agent can just send me feedback right away. And as I mentioned, you can use your favorite agent, whether that's Cloud Code, Codeex, Pi, Hermes,

1:25:05 Cursor with your existing subscriptions. All of that you can use inside of Cloud Room. Go to cloudroom.dev or click the first link below video and give it a shot. Um, yeah, I'm highly confident

1:25:16 that it's the best AD in the world, the best AI cooling tool in the world, free and open source, finally on Windows. You get free cloud agents. Again, the only tool in the world that gives you cloud

1:25:26 agents, even on the free plan for any harness, the only in the world. So, give it a shot. And uh yeah, cloud.dev, first link video. First link below the video. With that being said, thank you guys for

1:25:37 watching and I wish you a wonderful, productive week. See you.