YT YouTube2Text

YouTube2Text → Transcripts → My Agentic Engineering Workflow (after 6,775 sessions)

My Agentic Engineering Workflow (after 6,775 sessions) — Transcript

by David Ondrej · 11,345 words · 1,572 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00My name is David Andre and over the past
0:01three years, I've spent well over 2,000
0:04hours coding with AI. And I've also
0:06personally interviewed some of the most
0:08legit and productive people in the space
0:11of Agentic Engineering. And in this
0:13video, I'm going to show you my full
0:15Aentic engineering setup as it currently
0:17stands, holding nothing back. Starting
0:19with the interface, right? This is how
0:20you interact with agents. So currently,
0:23my main interface is BB. This is an open
0:25source project. It's completely free and
0:27it allows you to use any subscription,
0:30any agent, any model inside of a single
0:32interface. Right? So we can see we have
0:34CEX, cloud code, pi, cursor CLI, open
0:37code, uh grot and hermes all within the
0:40same UI. The problem with apps like you
0:42know codeex or uh cursor is that they
0:45only allow their own models, their own
0:47subscription and you cannot utilize the
0:49most out of your subscriptions which is
0:51absolutely essential in 2026. Now inside
0:53of BB there's a lot of different useful
0:54things. You can see that I have
0:55different projects on the side. I can
0:57easily close them. I can easily open
0:59them, reorganize them if I wanted to.
1:01There's also agents and you can launch
1:03sub agents. So, for example, here I told
1:06it to run the total review skill. I'm
1:08going to go a bit more into detail on my
1:10skills later on. But basically, what it
1:11did is it launched these two sub agents
1:13and uh yeah, I didn't have to write that
1:16prompt. I didn't have to manage this sub
1:17agent. The parent manager agent launched
1:20them and you can easily click into them.
1:22It's very nice. all the features you
1:24like from uh you know cursor or codex
1:27apps which again those are two great
1:28apps don't get me wrong they are inside
1:30of BB and the benefit is that BB is
1:32fully open source so you can fork it you
1:34can look into side of the code and again
1:36it's free to run that's why I use it but
1:38it's not the only thing in my agentic
1:40engineering setup in terms of the
1:41interfaces another one is CMAX cm is a
1:44great way to have different projects uh
1:47you can see here for example I have a
1:48bunch of codexes so I will split this
1:51thing five different codexes how they
1:52behave when they lose internet access
1:54and you can do that easily because when
1:55you launch a new workspace inside of
1:57CMAX you can just uh divide the screen
1:59like this just like you would do in
2:01T-Max and that's why the similar name
2:04it's a terminal right so each of these
2:05are terminal panes uh you can launch
2:07different harness there's also built-in
2:09browser CMAX is great the the issue with
2:11CMAX is where it breaks is once you have
2:14a lot of different agents a lot of
2:16different workspaces the left sidebar
2:17really is not the right primitive it's
2:19not the right interface it it's really
2:21hard to work add big projects inside of
2:24Cmax. It's fine if you have like a
2:26couple of things going on and if you're
2:27using couple of agents, but anytime you
2:30want to do serious energetic engineering
2:32work, CMAX is not going to work. Another
2:34thing I'm using is Ghost T as my
2:36terminal because it's very fast and it's
2:38native and inside of Ghost Ty, you can
2:40run her which is basically like T-Max
2:43but for agents. It's a a backend runtime
2:46for agents. Herder is really good. It's
2:48lightweight. The issues I had with for
2:50example cursor app is that it was
2:52breaking. It was just crashing and it
2:53was running like five plus agents in
2:55work trees. More on work trees in a bit.
2:57But herder is very minimal, very
2:59lightweight. It lives in the terminal
3:01and allows you to um again start new
3:03agents and new sessions very easily and
3:06you know cd into folders and build stuff
3:08very quickly, very easily. And when the
3:10agents finish their state is clocked on
3:11the left right. So you can see which
3:13agent is done, which is idle, which is
3:14blocked and which is running which is
3:16essential. So for example here if I
3:18start like codex and I can say hi you
3:20can see he's working right he's moved to
3:21the bottom he's working and you don't
3:23have to pay attention to him and now
3:24he's idle right so this is very
3:26important tracking the states of AI
3:27agents and my prediction is that in 3 to
3:306 months as agentic engineering grows as
3:32a field we're going to see this become
3:34more and more important because you're
3:35not just going to be talking to a single
3:36agent you're going to be talking to a
3:38manager agent that manages a lot of
3:39different worker agents because again
3:42human attention is fundamentally limited
3:44that's why having a good interface is
3:46essential Right. Another thing is coral.
3:48This is something I developed myself. It
3:50had multiple versions. This one is fork
3:52of zeron. But basically the way this
3:54works is that instead of randomly
3:56switching between agents whenever they
3:57finish, which is you know how like for
3:59example inside of herder it works.
4:01There's no real order. What I realized
4:03is that like every agent has a different
4:05importance, right? Different priority.
4:06So when you have like a agent that's P1,
4:08when he finishes, he should go to the
4:09top. And it should never you should
4:11never like respond to an agent that's P4
4:14when a P1 agent has finished running. So
4:16this is what I call coral. I still need
4:18to like open source and finish it. We'll
4:20see if it has any reception. But this is
4:22in terms of interfaces. That's what my
4:24agentic engineering setup looks like.
4:25Now I already touched on this briefly
4:27but the next thing I need to talk about
4:28is models and subscription. Right? How
4:30do you get the most value out of your
4:32money? So these are the four main
4:34subscriptions right now. And again, this
4:35could be completely different two months
4:36from now, but right now the best deal
4:39completely is open code go. Okay, $10
4:42and it gives you Kimik K3. It gives you
4:45Gro 4.6, GLM 5.3, Deepseek V4 Pro.
4:50Insane $10. Really good, but it doesn't
4:52have the best models like Fable and GB
4:545.6.
4:56And the limits are kind of small, right?
4:58So once you get open code go if you can
5:00afford more money you need to buy
5:02another subscription
5:04$20 you're going to start with a $20
5:05subscription probably start with chbt1
5:08or cloth you know if you really have a
5:09strong preference or cursor if you want
5:11to have more models the issue is with
5:14cursor you don't get as large of a
5:15subsidy right these two subscriptions
5:17will give you the best deal in terms of
5:20like usage but they only have their own
5:22models right the this one only has JGBT
5:25models this one only has anropic models
5:27so Again, if you have $30 to spend, go
5:30with Open Code, go with CHBT Codex. If
5:32you have $50 to spend, add in Cloth Code
5:35$20 subscription on top of that. If you
5:37have $70 to spend per month, add in
5:40Cursor's $20 a month subscription and
5:42you have the lowest tier of all
5:43subscriptions. Now, once you have more
5:45budget, let's say you have $110 per
5:47month, you want to go with one of the
5:49big plans, right? Because these are the
5:51ones that give you by far the best deal.
5:54So, CHGBD Codex is going to give you a
5:56better deal than CLOT. It is what it is.
5:58You know, OpenAI has more compute.
5:59They're willing to subsidize it more.
6:01So, for the foreseeable future, if you
6:04go with a CH GB subscription, doesn't
6:05matter if it's 100 or 200, you're going
6:07to get a better deal than Cloud Code.
6:10Now, if you can, you should afford both,
6:12right? If you have like $210 a month to
6:14spend, you should just get this one,
6:16this one, and this one. Cursor is very
6:18underrated. A lot of people are don't
6:20realizing that Cursor aka Grock is going
6:23to become a great subscription because
6:25of SpaceX acquisition, right? SpaceX AI,
6:28Elon Musk's company, has uh loads and
6:31loads of compute. They have so many
6:33GPUs, which means they're allowed to
6:35play the game. They can play the game of
6:38subsidizing. Now, it's not true yet. You
6:40know, cursor subscription is not yet as
6:43subsidized as codeex or cloth code. So,
6:46keep that in mind. But very easy
6:48prediction that I'm confident in. In 2,
6:503, 4 months, we're going to see the
6:52cursor subscription become a great deal.
6:54Already with Grock 4.6, it's a great
6:56deal. Gro 4.7 is all around the corner.
6:58This is going to become an insanely
6:59valuable deal. I I predict. But right
7:01now, these two subscriptions are better.
7:03And you know, if you can afford it, just
7:05get the 200s because those are 20x,
7:07right? No matter what, do not pay API
7:10pricing. Do not pay API pricing. This is
7:12the worst deal. Uh because, you know,
7:14there is no subsidy. There is no deal.
7:16There is no uh like more tokens for
7:18less. You want to get as many tokens as
7:20possible for as little as money as
7:21possible. That's why subscriptions are
7:23essential and u yeah depending on your
7:26budget. Just combine these four however
7:28it makes sense for you. Next up, cloud
7:31agents. It's pretty inevitable that
7:32cloud agents are the future. I think
7:34this is obvious to anybody who's been
7:35paying attention to what's happening in
7:37agentic engineering. You can see the
7:39moves of cursor, AMP, Devon, Codex, all
7:43of these companies are just going all in
7:45on cloud agents. And by the way, the
7:47proof is just this graph, right? This is
7:50internal 7-day rolling share of merged
7:53PRs from cloud agents at cursor. And you
7:55can see that the start of this year, it
7:57was like around 10 15%.
8:00Now in uh you know, second half of 2026,
8:04it's the majority. It's approaching near
8:0660%.
8:08which means most of the work done at
8:10cursor inside of cursor and this is not
8:13just like you know automations that run
8:16a PR and it doesn't get merged this is
8:18merged PRs so it's the stuff that
8:19actually gets used most of it comes from
8:22cloud agents okay so what is a cloud
8:24agent well all of the things that like
8:26I'm showing you for example here in BB
8:28these things are running locally these
8:29things are running on my computer right
8:32they're local local agents running on my
8:34machine the issue with this is that it's
8:37not scalable You cannot run hundreds of
8:39agents at once because your computer
8:41will probably not handle it. All it
8:42takes is couple of agents to decide to
8:44run database tests at the same time or
8:46to run your entire test suite at once
8:49and it's going to use up most of the
8:50resources of a computer. Another thing
8:52is isolated environments. Another thing
8:54is persistent sessions uh durable access
8:56to internet or electricity. You know
8:58when you close your laptop you lose your
9:00sessions um when you lose even when you
9:02lose internet for like a couple minutes
9:04um the harnesses cannot recover. So
9:06cloud agents are absolutely the future.
9:09Not 100% obviously there's still going
9:10to be use case for running agents
9:13locally on your computer but more and
9:16more of the AI that's running and exists
9:19is moving to the cloud. I mean just look
9:21at Grogbot, right? Grogbot is again a
9:23hugely successful product. It's blowing
9:24up and it's just another way of doing
9:26cloud agents. The problem however that I
9:29have with these solutions is that they
9:32have insane level of ecosystem blocking.
9:34Right? When you set up any of these
9:35cloud agents and you set up the
9:36environments, add all of your secrets
9:38and environment variables
9:41that is first of all a lot of hours to
9:42set it up correctly and second of all
9:44like you just lock yourself in. You
9:46know, you have all your agent sessions
9:48there. You're using their subscription
9:50or their pricing and you give them all
9:52your context and they can train models
9:53on it. And even if they don't train
9:55models, that's the thing. There's so
9:57many other ways to use your data. They
9:58can look at it, they can analyze it,
9:59they can compare it even if they don't
10:02train on it, right? So the biggest
10:04problem is ecosystem lockin. Like once
10:06you set up any of these, if you migrate
10:07your whole company to them, switching
10:10away is going to be such a pain and so
10:11expensive. So what is the solution?
10:14Solution is to have your own server,
10:15right? It's not that scary. It's
10:17actually very simple. I'm going to show
10:18you in a second. And on that server, you
10:20run her and you SSH into it. Herder
10:22gives you persistent agent sessions and
10:24you can have many agents running at
10:26once. And SSH allows you to connect to
10:28your server in a easy way. Okay, you can
10:30do it from your phone. You can do it
10:31from laptop. Anything that supports SSH
10:34and you can literally achieve the 8020
10:36for just a couple of dollars. So, let me
10:38show you how to actually set this up so
10:40you can benefit from this cloud agent
10:42revolution for a fraction of the cost
10:45without locking you into an ecosystem
10:47that you might not want to be in. So,
10:50the first thing you need to build your
10:51own cloud agents is a server. And the
10:53easiest way is to just buy a VPS. So, I
10:55think the best option is Hostinger. I've
10:57been using them for years now. All of my
10:59VPS's and everyone on my team that has a
11:01VPS uses Hostinger. Not only because
11:04it's one of the most affordable options,
11:06but because it's so easy to set up. Let
11:08me just show you. So, when you get to
11:09the landing page, which by the way, I'm
11:10going to link that below as the first
11:11link. Just select the plan. KVF2 plan is
11:14enough. But if you prefer to go more,
11:16like, you know, if you want to have
11:18hundreds of agents running at the same
11:19VPS, you can just go with one of the
11:21more powerful servers. I'm going to go
11:23with KVM2 here. So, click on choose
11:25plan. This is going to take you to the
11:27Hostinger card. Now, for the period, go
11:29with either 12 months or 24 months. I
11:31would recommend going with 24 months
11:32because you get the biggest savings, the
11:34biggest deal. And again, every month
11:36we're running more and more EI agents,
11:37so you might as well get early while
11:39other people take another 6 months to
11:41realize this. Now, again, this gives you
11:42the biggest cost savings. But if you
11:44want to save additional 10% off, go to
11:47the right, click on have a coupon code,
11:48and type in code David to save another
11:5110% off on the one year or two-year
11:53plan. Then on the left, we don't need
11:55the Noxious credits. You can uncheck
11:57that. We don't need this. Keep all of
11:59these unchecked, but select a server
12:00location that's close to you. So, right
12:02now, I'm in Warsaw, Poland. So, Germany
12:05is pretty close. So, select that. But
12:06then at the bottom, you need to select a
12:08location. So, just click on that and see
12:09like what's the latency to where you
12:11are. So, you can see Germany's at the
12:12lowest latency. So, I'm just going to go
12:14with that and then click on continue.
12:16Next, you need to log into your
12:17Hostinger account or if you don't have
12:18one, just register. Takes like 20
12:21seconds. It's super easy. And then fill
12:23out your card and billing details to
12:24complete the purchase. Okay. Once you
12:26purchase the VPS, it will look like
12:28this. You'll get redirected to the
12:29hostinger panel. Uh on the left, you can
12:31always go to VPS and click on the one
12:33you just purchased. So, uh this is your
12:35IP address, IPv4. This is your root
12:38access and uh the password. Hopefully,
12:40you saved that during setup. If you
12:41haven't, just click on reset password,
12:43create a new one, and save it to your
12:45password manager. Then, once you saved
12:46your root password, what I actually
12:48would do is I would use an agent to help
12:50you set up your environment for cloud
12:52agents. So, for example, we can use CMAX
12:54here. Okay. So I went into root level.
12:57So again CD you can change directory go
12:59to root level of your computer to you
13:02know or go into a specific folder that's
13:04up to you but the main thing is
13:05launching a coding agent right you can
13:07use cloth code you can use uh codeex I'm
13:10going to use cursor CLI because you can
13:12use any model here and for this I think
13:14Grock 4.6 is good uh we don't even need
13:17extra fast I'm going to select high and
13:19fast um Grog 4.6 ICS is surprisingly
13:22great to work with. So, we're going to
13:24use it here. I'm going to need your help
13:26to SSH into a VPS and set up some
13:28dependencies here. Always answer in
13:30plain English. Be very concise. So, this
13:33is what I call presending, right?
13:34Sometimes it's good to like present a
13:36prompt to tell it like what is this
13:38session about? And uh okay, so it's
13:41actually in following my skill, which by
13:43the way, you can get from my skills
13:44repo. That's also going to be linked
13:46below completely free. So, I already
13:47have a skill that teaches the agents how
13:49to use um SSH, how to use hosting VPSs,
13:53but I'm going to tell it, hold up for
13:55now. This is a brand new VPS I set up.
13:57Don't use any skills yet. Don't uh rush
14:00anywhere, okay? I'm going to tell you
14:02what to do. So, I'm using Super Whisper
14:04here to dictate because it's way faster
14:05than typing. Even though I type very
14:07fast, if you want to know how I type
14:09fast, just go to mike type.com and, you
14:12know, do some typing. You can measure
14:14it. But even if you type 150 words per
14:16second, you can speak at 300. If you are
14:19really fast, you can probably speak at
14:20least 250, right? So most of you are
14:22slow typers. You type at like 40 or 50
14:24words per minute. You can speak three to
14:27four times faster instantly. So just get
14:29a voice transcription tool. You know,
14:31there's a Glo from Jack Roberts. That's
14:33a good one. I use Super Whisper. There's
14:36a Whisper Flow. It doesn't really matter
14:38that much. Okay? Use a voice
14:40transcription tool. Um, sometimes you
14:42cannot use it if you're in public.
14:44that's fine. But if you can use it, use
14:46it. It's super OP and it will allow you
14:48to send prompts way faster. And that's
14:50what you're going to see me watching in
14:51this video when I'm able to speak into
14:53my microphone and get the prompts out
14:55super quickly. It's because I use super
14:57whisper. So anyways, we have the agent
14:59primed here and we have the other pane.
15:01So I'm going to actually tell it use
15:02Cmax. Okay. So, I'm going to tag the
15:04skill. Use the skill to find the other
15:08empty pane uh in this uh CMAX workspace
15:11where I just ran the command pwd. Let me
15:14know if you can find this pane. This is
15:16where I want you to run all future
15:18terminal commands once I tell you what
15:21VPS to SSH into. This is why I like to
15:24use CMAX for these types of setups is
15:26because uh first of all, it's like
15:27straight in the terminal, right? You can
15:29see everything that's happening. If you
15:31use like claw the desktop app or codex
15:34app, it hides a lot of the details. And
15:36you know, it's good for beginners. It's
15:38good for nontechnical people. But if you
15:40want to be a serious agentic engineer,
15:42you cannot be scared of the terminal.
15:44The terminal is your friend. It's just a
15:46way to access the computer, right?
15:48There's nothing to be scared about. So
15:49it found it. Okay. So that's good. It
15:52found the pane. So now it can interact
15:53it. So like send a test command in
15:56there.
15:58I just want to test if the agent can
15:59interact with this pane and because of
16:01this skill right of the CMAX skill which
16:03again it's in my skills repo. Uh it can
16:06and it sent echo test. This is very
16:09important because I'm going to use this
16:10agent to do everything. I'm going to
16:12give him the credentials to the VPS and
16:14I'm going to have him set up the
16:15environment and I'm just going to talk
16:16in plain English. This is the alpha
16:18guys. You can use the agents for so much
16:20more than you're currently using it. Let
16:21me just show you. Okay. Okay. So, let me
16:23switch back to Hostinger. I'm going to
16:24copy the root access command here. I'm
16:26going to go to the right pane, paste
16:28that in. And for the credential, like,
16:30you know, pasting the IP and the
16:31password. You probably want to do that
16:33yourself. If you're repeating it long
16:34enough, you can store it in a secret uh
16:37env.local
16:39and mention only the location of the
16:41environment file in your skills. Never
16:43put credentials or enviables into skills
16:45or into prompts. Okay, it's a very bad
16:47practice. But anyways, now it's asking
16:48for the password. So, I'm going to fill
16:49in the root password. There it is. And
16:51now we're SSH. So, I'm going to speak.
16:55Okay, check it again. Check the pane
16:57again. I just SSH to the VPS. Do you see
17:00it? If so, run a few test commands to
17:02learn more about that virtual private
17:04server. And now we're going to see the
17:06speed gains of using an AI agent, right?
17:08So even though like even if you're a
17:12professional developer or a backet
17:13engineer, DevOps expert, you can only
17:15read so fast, the agents can read
17:18insanely fast and they can run write
17:20code and terminal commands like super
17:22quickly. So what the agent can do is
17:24basically analyze the VPS and learn
17:25everything about it in a matter of a few
17:27seconds. Right? So that's what it
17:29learned. Ubuntu blah blah blah to vCPU
17:32cores about 8 GB of RAM. Okay. So great.
17:35Great. Exactly. So now I want you to set
17:37up the following
17:40dev environment on the on that VPS.
17:42Right. So we need to install her. Also
17:44make sure there is Node.js. Also make
17:46sure there is Python 3. Make sure
17:48there's Git. Start by installing all of
17:51these. And instead of me like having to
17:53go and visit these sites and copy their
17:54installer commands, the agent can just
17:56make sure that we get the latest
17:57version, the latest stable releases of
17:59all of these dependencies to set up our
18:01cloud agents environment on this VPS.
18:03And again, a single hosting VPS like
18:05this can handle many different agents.
18:07You can literally have a Hermes agent
18:08running with your personal stuff. You
18:11can have another open cloud there and
18:12then you can have all of your coding
18:14agents, right? Cloodes, codexes,
18:16cursors, droid, prime agent, whatever
18:18you're using, it doesn't matter. And if
18:20you ever need to upgrade, it's very
18:21easy. You can just do it from your
18:22hostinger panel, which takes, you know,
18:24a couple of clicks. Much easier than
18:25upgrading your home server or your
18:28entire, you know, setup at home, buying
18:30a new laptop, migrating everything over,
18:32that is a lot more expensive and a lot
18:34more inconvenient. And again, you need
18:36some cloud setup. You need some cloud
18:37environment set up where if you have a
18:39long running task and you want to start
18:41it and let's say you need to go to a
18:43cafe or you need to close your laptop or
18:45you are in a place with a buggy internet
18:47you cannot run it locally on your
18:49computer. So you use cloudations for
18:51that and with the your own VPS you can
18:53customize it however you want. And the
18:55main thing I want to get across is that
18:57you don't need to be an expert in terms
18:58of VPS's. You don't need to be an expert
19:00in terms of DevOps Linux none of that.
19:02Just talk to your agent in plain
19:04English. Look at what I'm doing. I'm
19:05talking to Grock 4.6 in plain English
19:07and he's doing all of this for me. So,
19:09whatever you want, whatever cloud setup
19:12you want, you know, connectors, whatever
19:14harness, whatever interface, however you
19:16want it to work with, how you want it to
19:17to interact, to work, chronops, whatever
19:21you want to set up, describe it in plain
19:23English. Tell it to give you free
19:25options and just walk through the main
19:26decisions. Educate yourself. Focus on
19:29main choices, but forget the syntax.
19:31Forget specific terminal commands. The
19:33agent can do it. So literally all you
19:35need is CMAX and a coding agent running
19:38on the left. on the right. You just do a
19:39terminal into VPS and you can set up any
19:41type of cloud agent setup and you can
19:43completely avoid all of the ecosystem
19:46lockin that comes from these companies
19:48which they have a massive incentive to
19:50move you to their own cloud agents
19:52platform right why because it's very low
19:55churn because once you set up you know
19:57any of these and I get I don't want to
19:58hit on a specific company but any of
20:00these once you set them up it's nearly
20:03impossible to switch away so I would
20:05recommend doing that and instead get
20:07your own VPS PS, set up herder, set up
20:09SSH, and just run your own cloud agents.
20:12It's not that difficult. We're like,
20:13we're literally in the second half of
20:142026. AI agents are super powerful. Just
20:18describe what you want in plain English
20:19and have them build it for you. In fact,
20:22it's done now. So, we can literally say
20:24uh start herder or we can just run that
20:26command oursel again. I want to show you
20:28the power that really agent can do
20:30anything. So, it's going to read the
20:31state of the interface and now it
20:33started the herder. You can see that we
20:34have it here. Let me resize that a bit
20:36more. So now here we can like for
20:39example launch PI agent. We haven't
20:41installed PI agent. So I say now install
20:44PI agent use deep API to learn what is
20:48the official installer. Boom. And that's
20:51another OP thing. It's like you really
20:53need one agent and then use that agent
20:55to compound and set up other ones,
20:56right? So, for example, for Pi, uh, it
20:59can find an installer, it can install
21:01it, it can set it up, it can probably
21:02find an open router API key on my
21:05machine and copy it and can do the full
21:07PI agent setup for me on this fresh VPS
21:10just by me telling it, right? So,
21:12really, if you have a computer, you need
21:14to fight to set up the first agent,
21:16right? No, I don't care what it takes. I
21:18don't care if you need to cancel a
21:19meeting. Sit down, set up the first
21:21agent on the computer.
21:24After that, everything else is so much
21:25easier. And yeah, look how easy it was
21:27to set up herder. We have the spaces, we
21:29have agents. Obviously, um the agents
21:31are not installed yet. Can try cloth
21:33code. Okay, it's installing pi now, so
21:36I'm not going to interrupt it. But yeah,
21:38that's how fast it was to install her.
21:40It's running on the VPS. Really just
21:41talking to your agent and describing
21:42what you want is underrated. All you
21:44need is work ethic initiative and just
21:46push through the friction. That's it.
21:47Now, while the Pi is getting installed,
21:49let me continue with my agentic
21:50engineering setup and talk about
21:52harnesses because this is what we
21:53already touched on. So, let me go deeper
21:54on that. First of all, Pi Agent, the
21:56goat, you know, the simplest, most
21:58minimal hardness out there. If you want
22:00to get it, pi.dev. Again, it's open
22:01source. It's completely free. Just
22:03download it. Um, I had Mario on the
22:05podcast. And also, Armen is releasing,
22:08the episode of Armen is going to be
22:09released in the next week or so, so stay
22:11tuned for that. Make sure to subscribe
22:13if you don't want to miss that. I've
22:14been really doubling down on the podcast
22:16interviews, interviewing some of the
22:18most impressive people in the AI space.
22:20So if you want to make sure that you see
22:22all of them and you didn't don't miss
22:23any of them, please subscribe. It's
22:25completely free and it takes two
22:26seconds. So go below the video and click
22:28subscribe. Okay. So PI agent again this
22:30is what OpenClaw is built upon. It is a
22:32super minimal hardness with just four
22:33tools and yeah it always runs in YOLO
22:36mode. It supports any model any
22:38provider. Very elegant very easy to use.
22:41Super configurable configurable and
22:43customizable. And this is why it's
22:45becoming one of the most popular
22:46harnesses out there and why a lot of
22:48people built on top of Pi. So absolutely
22:50non-negotiable. You need to be using Pi.
22:53You need to have it installed. And
22:54that's why I put it as the first harness
22:56on this VPS. Seems like it is installed.
22:58So now I say, "Okay, good. Now go
23:00through the setup. Um, go through it
23:03fully yourself. And if you need an open
23:05router API key, just analyze my MacBook
23:08and find one and then use that same one
23:10on the VPS to get Pi running until you
23:13get a model. Let's go with GPD 5.6.
23:17until you get that model running inside
23:19of the PI agent with open router. Keep
23:21going through the setup. Uh you have my
23:23permission.
23:24Boom. And literally Grog will just do
23:27all of it. And again, it doesn't matter.
23:29I'm using cursor CLI, but you can use
23:31cloth code, you can use codex, whatever
23:33you want to use, just use it and it it
23:36will do the full setup for you. So,
23:38while it's finishing the PI agent
23:39walkthrough on boarding, what is the
23:41other harness? Cursi. This is what I'm
23:44using. very underrated harness uh
23:46because you can use all the models right
23:48inside of cursor as long as you have
23:49some subscription. You can use grock,
23:51you can use GPT models will actually
23:53open AAI wants to remove them but that's
23:55another story. You can use enthropic
23:57models. You can use Kim K3. So many
24:00different models. And this harness is
24:01actually good. Like if you aren't using
24:04it, you can like easily tag skills, you
24:06know, you can presend messages.
24:09Even close code doesn't have the ability
24:10to preset messages. So CLI very
24:13underrated harness. Next up, we have
24:16self-learning harnesses, right? So this
24:17is for stuff where you're not sure what
24:19to do. If you're starting some task
24:20where you're like, hm, I don't really
24:22know what this should look like. I don't
24:24really know what I'm doing here. You
24:26need to be using a self-improving
24:27harness. And the two main ones are
24:29Hermes agent. Of course, Hermes is the
24:31OG when it comes to self-improvement and
24:33self-learning harness. Prime Agent is a
24:35bit newer. Um is a very different
24:37approach compared to Hermes. Both are
24:39promising. Just make sure you use a
24:42selfimproving like choose one. Okay?
24:44Test both. Choose one of them. If you
24:47are doing a task where you don't know
24:48what you're doing and you have a lot of
24:50uncertainty, a lot of confusion, a lot
24:52of figuring out to do, use a
24:54self-improving harness because this
24:56creates skills and improves with you
24:57over time. If you're doing something
24:59specific like I'm doing inside of Cmax
25:00to set up or like if you're developing
25:02software or you don't want random skills
25:04created, then yeah, just stick to codex
25:06or you know, pi or or cloth code or
25:09cursor. But if you're doing something
25:11new, use either Hermes or Prime. Next
25:14up, the classic cloth code or codex. The
25:17reason I put it like this is because um
25:19so for me, all I need to do is type in
25:21CC and this will launch cloth code with
25:23dangerously bypass permissions. A lot of
25:26people are doing u unnecessary typing by
25:30doing clo-dangerously
25:33skip permissions. Like there's no way
25:34you're typing this daily. Okay,
25:36extremely slow, extremely unefficient.
25:38just create a global alias like CC
25:41launches cloth code with dangerous
25:43script permissions and I have the same
25:44thing with codeex just type in CX
25:46launches codex in yolo mode so I have to
25:48type in the parameter the argument every
25:51single time right it's very inefficient
25:53very slow that's like the half of the
25:56agentic engineer it's like how can you
25:58work faster how can you get more
26:00done in the same amount of time and
26:02creating these global aliases for the
26:05long terminal commands you run often is
26:07one example So absolutely essential and
26:10yeah these are the harnesses that you
26:11should be using and that I'm using. Next
26:14up let's talk about skills right. So my
26:16skills repo went viral last month. A lot
26:19of famous people in the AI space
26:21retweeted it and posted that it's one of
26:23the most valuable repos. That's their
26:25words not mine. And I made a full video
26:26on this actually recently right here. So
26:29if you want to watch that just go ahead
26:30and watch that later. It's a 30-minute
26:32video just breaking down um eight skills
26:35from my skills repo that I absolutely
26:37use all the time. But specific to this
26:39video, you know, the agentic
26:40engineering, I picked out ones that are
26:42the most relevant to my AI coding or
26:44agentic engineering setup. The first one
26:46is total review. So this one you can
26:48access. Again, my skills repo is
26:50completely free, guys. This is, you
26:51know, this is I don't make any money
26:52from my skills repo. It's totally free,
26:54open source. Um feel free to get it.
26:56Feel free to fork it. This one if you go
26:58into skills folder
27:01and you go into agent orchestration you
27:03can see at the bottom is total review.
27:05What this does is that it runs two other
27:07skills GP review and fable review to
27:10just review any changes you're doing.
27:12Right? So if I go into VB and let's say
27:14here I just ran it. So this was a like a
27:17preparing this repo for public release
27:20and I just told it to review the changes
27:22with total review and instead of me
27:24saying like launch fable give it the
27:26context tell it to review everything
27:28like a senior developer. Why would I do
27:30that? Anything you repeat anything you
27:33do more than enough you should either
27:35turn it into a prompt. So for example
27:36look I have these uh keyboard shortcuts
27:38right? This is text replacements. So
27:40anything I'm doing like often enough I
27:42can just type instantly. So like answer
27:44in short in plain English. Now get to
27:45work and execute this plan fully and
27:47completely like a tennics engineer
27:48would. Make your previous answer simpler
27:50and shorter formatted in a nice readable
27:52markdown. Be very concise. All of these
27:54things I'm typing them instantly. Why?
27:56Because I got text text replacements.
27:58I'm super tryh hard. So I have them as
28:00raycast snippets. But you can start just
28:02by using the text replacements on Mac
28:04OS. It's perfectly fine. Or just use
28:07raycast if you want the fastest ones
28:09like I'm using right here. But look at
28:11these prompts, right? this prompt, it
28:12would take me like two minutes to speak
28:14out and I would probably make a mistake
28:16and it wouldn't be consistent every
28:18time. I just have it as a as a you know
28:21preset. Same thing with like make your
28:23answer short in plain English. Same
28:25thing with like stage files, write a
28:26clear comment, push to GitHub. Don't
28:28overthink this, right? These are the
28:29things I'm repeating. So if it's like a
28:30single step, you want to have a text
28:32replacement. If it's a more advanced
28:35workflow, you want to have a skill such
28:37as total review. This invokes two other
28:40skills which are called GPT review and
28:42fable review which again it's pretty
28:44self-explanatory reviews whatever code
28:46changes you just did with either GPD 5.6
28:49soul or fable 5.
28:52This is very important because you want
28:53to review your changes with the best
28:55models right but total review takes you
28:56to the next level. it runs both of them
28:58and then it decouples them into a single
29:01list. So it, you know, you're not like
29:04wasting tokens. And yeah, when anytime
29:06you do mediumsiz to large changes, you
29:09should just run total review on it to
29:11make sure it's all good. And especially
29:12if you run a different model. So for
29:14example, if we run something with uh if
29:17we do something with let's say Grog 4.6,
29:20you definitely want to review what
29:21changes Grog does with GBD 5.6 six soul
29:24or fable 5 because it's a totally
29:26different model and it will find it in a
29:29different way, right? Like imagine if
29:30you're doing something important, maybe
29:31you're submitting a job application or
29:33really something that matters in your
29:34life and you ask one smart friend to
29:37review it. Obviously, that's a good
29:38idea. This total like that's that's
29:41basically what like GPT review is or
29:42fable review. Total review is like you
29:44ask the all of your smartest friends to
29:48review it and then you tell them to like
29:50just put it in the WhatsApp chat and
29:52like only give you the decoupled most
29:54biggest issues with your job application
29:56in this example, right? And they would
29:58just give you the things that really
29:58matter. This is total review. Next
30:01skill. I mean I guys I can speak on my
30:03skills forever. Again I I made a full
30:04detailed video on it. just watch that if
30:06you want me to go more in depth or
30:08comment below if you want me to like
30:09really break down these skills in more
30:11detail because this is what I spend
30:12hundreds of hours doing. Okay, so I
30:14don't want to make this video 20 hours
30:16long. So I'm just going to give you the
30:178020. Next up, ask and build. This is
30:20one of the newer skills u that I
30:22absolutely love. I use it every single
30:24day. Let's see where we can find it.
30:26Okay, thinking and docs here. So inside
30:28of the thinking and docs subfolder, we
30:31can find it here.
30:33And this skill basically what it does is
30:35before you do any building you run like
30:38decisions right so let's say here I
30:39would say like
30:41I want to make this Linux and Windows
30:45compatible
30:47run ask then build on the main five
30:50decisions. So instead of you saying like
30:53make this Windows compatible it's the
30:55issue with that like okay it can work
30:57but it's going to do some decisions. is
30:59going to make some choices that are like
31:01important architectural choices or
31:03important software design choices that
31:06you might regret later. Okay, if you run
31:08this skill as build, you frontload a
31:11little bit of work extra for making your
31:13codebase a lot more scalable, avoiding a
31:15lot more tech debt and just you know
31:18building creative software, better
31:19software. So this is what I use
31:21literally all day every day ask net
31:23build before any changes. So you see
31:25like how complete is the first Linux
31:27Windows ship and now you have four
31:28options. Do you want it to be same
31:30product on Windows? Do you want it to be
31:31Linux gets the full product then Windows
31:33gets binary only? Like you get to make
31:37the decisions. AI models are great at
31:39coding. They're great at implementation
31:41following a plan. They don't have the
31:43taste. They don't have good judgment.
31:45They don't have good decision-m you as
31:47the human need to stay in charge of
31:48that. And this skill helps you exactly
31:51achieve that. Now before I go through
31:52more skills, let me check what happened
31:54here. So seems like uh Grog finished and
31:58we have a fully running PI agent. Let me
31:59send a prompt. Hi.
32:01Fully running installed PI agent with
32:04open router as the provider powered by
32:06GP5.6. We can even choose the reasoning
32:08effort inside of PI. It's one of the
32:10main keyboard shortcuts. Shift tab to
32:13cycle between the reasoning efforts. And
32:15yeah, this is running fully inside of
32:17our VPS. So if I kill this, if my
32:20computer loses access, if if my MacBook
32:22explodes, those agents will keep
32:24running. And just like that, we have our
32:26own cloud environment for agents. And
32:28it's so like so easy. Like literally, I
32:30can launch a new one here. Boom. We can
32:32see we have different agents uh
32:34different pi agents here. And here I can
32:36do model and we can do let's say fable,
32:38right? And we are running fable and this
32:40one is running u GB 5.6ole. Both are
32:43running in the cloud on my own VPS. It's
32:45a virtual private server. I have full
32:48root level access to the YUbuntu
32:50machine. And yeah, it really is this
32:53easy to set up your own custom
32:55environment for cloud agents. And again,
32:57if you want to do that yourself, just
32:58grab yourself a Hostinger VPS. The link
33:00is going to be below the video. It's
33:01going to be the first link in the
33:02description and use the code David to
33:04save another 10% off. Oh, and shout out
33:06to Hostinger for sponsoring this video.
33:08Now, let me go to my skills again. The
33:11next one is deep API. This is what I use
33:14for any deep research, any scraping. um
33:17anything related. So like let's say
33:18we're about to do a a big like let's say
33:20here it's a good good example. So here
33:22before building the Linux or the Windows
33:24installation I would do something like
33:26use the deep API skill to run eight fast
33:28web searches with DB API to figure out
33:31what is the best way to do this on Linux
33:32and Windows like the different technical
33:34options you the different best practices
33:36to do this and present me with the top
33:38three options answer in short in plain
33:40English. Right. So again, I just voice
33:42prompted that super quickly and it's
33:44going to use the deep API skill to
33:46search up do in this case I instructed
33:48it very specifically to do fast web
33:49searches but you can also tell it to
33:51scrape you know Twitter what's happening
33:52to scrape GitHub uh to do a deep
33:54research whatever you need to do like if
33:57you have a person you can tell it to
33:58like find three different ways of
34:00contacting that person you know like
34:03find his email find his WhatsApp phone
34:05number find his LinkedIn whatever like
34:08this skill is so versatile basically
34:10anything the agents are struggling to do
34:12which you know codex and cloth code they
34:15come with basic web search but they
34:17don't have any scraping they don't have
34:19any DB research and they get blocked
34:20easily so that's where DB API comes into
34:23place it's very OP and literally I use
34:25it all the time everyone on my team uses
34:26it um yeah that's one of the main skills
34:29I use next up is two skills is the push
34:31lock and guard hook these are a bit more
34:33boring but absolutely essential okay and
34:36you only set them up once so if we go
34:38into ops and setup We can see it here.
34:42Global agent guard rails. This is a
34:44skill that also it comes with a hook. So
34:46if we go into the root level and go into
34:48hooks, all of these are connected that
34:50skill. It just makes sure your agents
34:52never like wipe your disc. Make sure
34:54your agents never override history and
34:57do any of these like dangerous patterns.
34:59I don't know how people live without
35:01this. Obviously, it's kind of crazy. Uh
35:03like there's so many things that like
35:05your agents could do. Like you could
35:08say, look at this, right? like recursive
35:10deletes, uh dangerous GitHub CLI
35:13commands or Git commands
35:16piping and downloading stuff from the
35:17web, you know, without uh without
35:21checking like just dangerous stuff. Like
35:23look at this. All of these commands you
35:25would never want your agents to run or
35:28at least these are patterns. So
35:29accessing like different password
35:30managers, right? Like imagine you get
35:32prompt injected and some skill like
35:34tells your agent to print out password
35:36manager or do something to your password
35:37manager. You never want to do that. Um,
35:40this is this guard drill prevents that.
35:41So again, you can find this inside of
35:43the skills repo in the skills folder in
35:47uh ops and setup global agent
35:50guardrails. Just run this and set it up
35:52on your computer. Do not ever use yolo
35:55mode or dangerous escape permissions
35:57without this. You're just risking too
35:59much for no reason. And anytime you see
36:01on Twitter somebody getting his uh you
36:03know disc wiped or database hacked or
36:06stuff like that all they could have had
36:08is a simple hook pre-tool call hook such
36:11as I have here with the global agent
36:13guard rails that triggers before the
36:15agent does the tool call. So it cannot
36:17trigger because the hook runs before the
36:19tool call.
36:21Another [clears throat] one is push log.
36:22This one is for those of you who are
36:25advanced, those of you who are running a
36:26lot of different work trees. The issue
36:28is still when you're merging these work
36:30trees into the main branch and pushing
36:32into production, if you're running many
36:35agents like let's say 15 plus in
36:37parallel, it will inevitably happen when
36:38there's like two or three agents who
36:40want to push at the same time. And um
36:43you know it's going to clog your CI/CD.
36:46It can like cause issues with merge
36:49conflicts and stuff like that. You need
36:51some Q. You need some lock on pushing.
36:55And this is uh how I do that. So I have
36:58the skill pro push.
37:02Pro push. And actually I can just have
37:06an agent explain it. So let me just go
37:08here. Boom. Explain how the pro push
37:10skill works. So the agent is a bit
37:12verbose, but I can just prefire. Make
37:13your previous answer simpler and shorter
37:15and format it in a nice readable
37:16markdown. Very underrated guys. Don't
37:19read long confusing messages. Just tell
37:21him make it simpler, shorter. You need
37:23to value your time. Your input tokens
37:24are just as valuable as agents. So the
37:27problem many agents can already explain
37:29this. So it's a one OS system OS level
37:31lock. So on the whole operating system,
37:33a real kernel log flock only one agent
37:37may run the full ship. Everyone else
37:38waits. So it's merge, verify, push, CI,
37:41deploy, health check. Right? So yeah,
37:43this is how it works. Absolutely
37:45essential. I couldn't live without it.
37:47But if you're not running many different
37:48worksheets, you probably can survive
37:50without that one. But again, my skills
37:51repo is going to be linked below.
37:53Completely free. Don't install all of
37:55the skills. Just grab the ones you need.
37:57You know, don't bloat your agents with
37:58skills you're never going to use. Next
38:01up, talking about work trees, right? So
38:03work trees, you should avoid them for
38:04small projects. What are workshes if you
38:07don't know? Uh it's basically copy of
38:08your folder. So your primary checkout,
38:10like if we go here, uh actually, let's
38:13go here. So here we're on the main
38:14branch. This is fine. This is a small
38:16project reposing just to use internally.
38:18I'm probably going to open source it
38:20actually. I think it's already open
38:22sourced. Yeah, I think I already open
38:23sourced it. The crazy part is I have so
38:25many things uh happening. I should
38:27probably delete that's the that's the
38:29dangerous part of AI you know you can
38:30build anything but
38:34if like if you can build so much stuff
38:36everything becomes distraction right so
38:39yeah I think this project has been open
38:41source yes it's already open source
38:42under MIT license okay good so here the
38:45reason I'm showing you is I'm not using
38:46work trees why because it's a small
38:48project here it would be absolutely
38:50overkill to use work trees and this is
38:52another good thing about BB is that they
38:54have built-in work trees so it's very
38:55easy to just launch launch a new work
38:57tree and it actually remembers it unlike
38:59the cursor app which doesn't remember it
39:01and you have to do /work tree every time
39:03which is extremely annoying but if we go
39:05to deep API which is a much bigger
39:07repository you can see that when I click
39:10plus it remembers that on the repos sync
39:13I work locally on the main branch and it
39:15remembers that on the API I always want
39:17a new work tree based off of origin/main
39:21really good this is why BB is my main uh
39:24interface for a genetic engineering. But
39:28basically the way to explain it is that
39:30you don't have to worry about worksheets
39:31at all if you're running a small project
39:33and if you're running less than 15 plus
39:36agents in parallel on medium to large
39:39size projects where you're running
39:40really 20 30 plus agents at all times.
39:44There's no way to avoid it. You need to
39:45use work trees and they're essential and
39:47you're going to be in a lot of pain if
39:49you don't use work trees because again
39:50these agents are going to be
39:51conflicting. They're going to be
39:52reversing each other's changes. they're
39:54going to be, you know, building other
39:56stuff. They're going to be like fighting
39:58each other. It's it's going to be a
40:00mess. Okay. What a work tree allows you
40:02to do is basically copies the primary
40:04checkout into a separate folder. And
40:05then in that folder, it creates a new
40:07kit branch so the agents can really work
40:10in parallel completely isolated from
40:13each other and not fighting against each
40:16other. It's as simple as that. So work
40:18trees depending on whether it's a small
40:21project you don't need them avoid them
40:22just stay on a single branch if it's a
40:25bigger project you absolutely need to
40:27use them. All right next up I'm going to
40:28do a rapid fire on some topics that um I
40:31mean you guys can see I can talk about
40:33this for 20 hours right so I want to be
40:34respectful of your time and let's do
40:36rapid fire on some agentic engineering
40:38topics. So you need to deeply understand
40:41when to use each model, right? Uh it's
40:42not just like okay, I have subscription,
40:44I should use this model. Like if you're
40:46designing the plan, you should probably
40:49use Fable. Like if you're working on a
40:51new project, use Fable. It has like the
40:53most sparks of genius. If you're fixing
40:56a deep serious bug that really needs
40:58attention to detail, 5.6 soul, max
41:02reasoning effort. You can do / goal even
41:05you know just go through it put more
41:07more effort than any other model for
41:08default chatting for like your main
41:11chatting of the model like here use grog
41:134.6 like usually high or extra high
41:15really good really consistent model
41:17really predictable model almost the same
41:19intelligence as fable and 5.6 soul, but
41:22it's two times cheaper and two times
41:24faster. So, a great model. Front end
41:26design, obviously, Kim K3 is right now
41:29the best for front end. Um, yeah, I
41:31mean, you need to know when to use each
41:33model, right? And again, this only comes
41:35with you actually getting your hands
41:36dirty and using all of these models.
41:38Don't be one of these people that like
41:39is only using the COX app and therefore
41:42only using GPD models and never testing
41:44anthropic models or, you know, Deepseek
41:46models or GLM or Kimmy. Don't be that
41:49those person. Okay? When a new model
41:51comes, legit set aside a day. Like
41:53seriously, guys, anytime a major new
41:56model is released, set aside a day where
41:58you can only use that model at least a
42:01day to really get a good feel of it.
42:03Like don't listen to what people say on
42:05Twitter. Try it yourself. Use it for
42:07your own workflows. See how good it is.
42:09Next up,
42:11uh, so you need to know when to push
42:14into production directly and when to
42:16like run total review or GP review or
42:19fable review, right? Like I'm not
42:20running the SL total review skill on
42:22every change if it's a small to front
42:24end tweak. I'm I'm pushing to prod right
42:26away. And again, this comes with
42:27experience. This comes with building
42:28more project with uh just, you know,
42:31creating more software. But uh you
42:33cannot just waste tokens by like
42:35reviewing everything. And also you
42:36cannot do these recursive reviews. This
42:38is a huge mistake. People go from like
42:40either doing no reviews, which is a
42:42mistake, or just like review after
42:44review after review. These models will
42:46invent imaginary bugs that are not
42:48there. Okay? So, just do one check.
42:51Okay? If you build something with Fable,
42:53review it with GPT, fix the worst
42:55things, and push it. Don't do like Fable
42:59into GPT into Fable into GPT. You're
43:01never going to ship anything. And these
43:02models will hallucinate. If you tell him
43:04find the five biggest issues with this
43:06codebase, it's not going to say like
43:07listen, the codebase is already good
43:09enough. You should do something else.
43:10That's what a human, you know, great
43:12developer would say. It's going to be
43:14like, okay, here are the five issues.
43:16Even if the codebase is like perfectly
43:17fine, so don't do this endless reviews,
43:20but also you should do reviews, you
43:23know, don't do zero reviews either.
43:25That's also bad. We already talked about
43:27this presending. Okay. So, yeah,
43:28basically
43:30I mean there's a funny video that uh we
43:33can actually put it here u on the
43:35screen. It's basically one of the people
43:37on my team recorded. It's me like
43:39presending like six different messages
43:40of me like knowing what happens, right?
43:42So basically when you have like fable or
43:44like let's say here it's like you know
43:45present with a plan then you can say
43:47like implement the plan. Then you can
43:48say like run fable review on this. Then
43:51if it's too long you can say like make
43:52your answer simple and shorter. And you
43:54know you can kind of predict these. So
43:55like say like when was the last git
43:58commit. Okay.
44:00So I can I can present that say like
44:02give me a plan to review those changes
44:06properly. Okay. So I can do execute that
44:09plan and I can do run fable review and I
44:11know it's going to already be concise.
44:13So I can do previous shorter and say
44:15like run stop overthinking and tell me
44:18what really matters and say like now fix
44:21those things. See like I can easily like
44:25usually it's like one or two or three
44:27sometimes it can be five or six I know
44:30what the agents will do and I know that
44:31the sequences look like if I'm if this
44:34is like too constant what I can do is u
44:37turn that into like a broader skill
44:39right so obviously it depends on the
44:42situation not every time is like so easy
44:45and elegant where you can just present
44:48the next five steps but sometimes it
44:49just lines up where you know like what's
44:51going to happen and what needs to happen
44:53or you can just present, right? And
44:55presending messages is essential. You
44:57should never use a harness or a GUI that
44:59doesn't allow you to present. Uh that's
45:01why, you know,
45:03BB is very nice because you can actually
45:05preset stuff. Okay, next up we have sub
45:07agents. So sub agents are very like
45:10overused. I feel like a lot of people
45:11just burn their limits on sub agents. I
45:13use them when I'm in control, right? For
45:15example, here we showcased how you can
45:18run the GBT review and Fable review, but
45:21I don't really like to have my agents
45:24randomly launching other agents. I know
45:26it's the future, but you need to have
45:28control over this. Like sub agents are
45:30definitely the future, and it's not
45:31going to be named sub agents. Like, I
45:33like the term manager and worker. It's
45:35way way more clear which agent does
45:37what. But if I run Fable 5, like a lot
45:40of these apps, they like delegated to a
45:43trash model, right? Like Haiku 4.5. You
45:45should never use that or GBD 5.6 Luna. I
45:47mean that's a lot better than the Hiku
45:494.5, but still I want to select which
45:52model runs in the sub agent. Okay, I
45:55know what subscriptions I have. I know
45:57what limits I have. I don't want to have
45:59a cheap model implementing my changes on
46:02a you know code base that I value. So I
46:04have my own skills such as you know the
46:06total review and similar or like /b sub
46:09aents uh which includes all the
46:12information about how to launch
46:14any sub aents inside of BB and I only do
46:17it when I want it. I don't let my agents
46:21randomly delegate random stuff to sub
46:22agents. I I don't think we're there yet.
46:25in the future. Obviously, it's going to
46:27be the future where you're talking to a
46:29manager agent and he launches a bunch of
46:30workers, but you still need to design
46:32that system. You still need to be in
46:34charge of the rules and the permissions
46:36and like the conditions when a sub agent
46:38is launched. I don't want somebody at
46:40anopic designing that like cloth code is
46:42full of slop, right? There's a reason
46:44why Pi exists. Pi exists because cloth
46:47code is bloated and there's a bunch of
46:49trash and slop and extra stuff that's
46:52wasting your tokens and and and changing
46:54your system underneath you. You don't
46:57want that. I don't want some guy at at
46:59Enthropic or OpenAI deciding when a sub
47:02agent is ran and which model is ran. If
47:04if you're happy with that, you're not a
47:06serious agentic engineer.
47:08Uh so yeah, you should be in charge of
47:10your sub agents. You should develop
47:12skills and rules and practices when sub
47:15agents should be started instead of just
47:17uh letting the default harnesses choose
47:19that for you. Next up we have ADRs. So
47:21basically this is a way to document
47:23decisions inside of a codebase so that
47:25future agents and humans can know why is
47:28it built this way. So I have it /docadr.
47:33Basically this is like one of the first
47:34folders I create. So here is a beautiful
47:36example repos. It's a small project and
47:39I have four decisions here like
47:41autonomous continuous operation secret
47:42guard file file name only default branch
47:45only homebrew cask distribution so
47:47whatever are the core architectural
47:49decisions in any project I document it
47:52inside of /doc/adr
47:54so that any future agents I can just
47:56point here and they instantly understand
47:58all of the decisions I made and anytime
48:01you want to change a decision in the
48:02future it's all documented the reasoning
48:04why you made that decision what was the
48:06state of the project at that time.
48:08Really essential like if you're not
48:09doing this, you should start doing this
48:11because the main thing is saving your
48:14context inside of a codebase so that
48:16future people or agents working in that
48:19codebase know how it came to be, how it
48:22evolved over time, why are there
48:24different decisions?
48:26Because some of the things can be read
48:28from code but not everything. And the
48:31the stuff that cannot be easily read
48:33from code should be documented as clear
48:35markdown files inside of this format. We
48:37already talked about that turning
48:38repeating pros patterns into skills or
48:41removing tests. Guys, what this is crazy
48:45especially if you use like I'll show you
48:47if you use soul gb 516 soul this model
48:51it adds so many tests even on the
48:54smallest repos like
48:56are there any tests on this project
49:00even on the smallest repos where it
49:02doesn't make sense these models will
49:03keep adding tests unit tests integration
49:05tests database tests component tests
49:09people at openai they just went crazy
49:11and They made this model always add new
49:13tests even when it makes no sense. So
49:16what you need to do is tell it like you
49:17almost need to tell the model to not add
49:19tests so that you land at the perfect
49:21amount. Okay, think about this. This is
49:23like advanced agentic engineering. If
49:26you don't tell the model to add tests,
49:27it like adds so much. It adds like
49:30insane amount. If you tell it don't add
49:32tests, it's still going to add some and
49:35then you land at the right amount. So
49:37maybe this will change with the next
49:39generation of models. Right now these
49:41agents just bloat your repo with tests.
49:43They add hundreds of tests even for
49:46stuff where it makes no sense. So yes,
49:49adding tests, unit tests, you know, end
49:52to end tests, whatever makes sense. When
49:54it makes sense, you should do it.
49:58But right now, I'm telling you, if you
49:59don't put some check on these agents and
50:01you tell them only add new tests when it
50:03truly makes sense. If you don't give
50:05them instructions like that, they're
50:08going to completely bloat your repo with
50:09pointless tests. Text snippets. We
50:11talked about that. Oh, this is huge.
50:13This is basically datadriven
50:15development, right? User like usage
50:18driven development. Maybe I should
50:19describe it like that. Any product that
50:21has real usage. You should give your
50:23agents read only access to production
50:25database. Actually, I have a skill for
50:26this pro uh read pro database. Right. I
50:31created a new like postgress role where
50:33these agents can have a read only role.
50:35So they cannot up like don't give
50:38your agents full access to database.
50:40Terrible mistake. You will regret it.
50:43You know maybe like 99.9% of the time
50:45it's fine. But all it takes is like one
50:49issue and it's like deletes a user or
50:51wipes a table some irre irreversible
50:54change and you're going to regret it. So
50:58don't give your agents right access to
51:00the database. But not giving it any
51:03access is also a mistake. Once you have
51:04some users, you can analyze your your
51:07data so much faster with an agent. See
51:09like what are the most popular things
51:11and features. What are the power users
51:13doing that the users that churn often
51:15are not doing? You know like what are
51:17the most common types of errors that are
51:20happening in production? You would have
51:22to go into that data look at it. It
51:24would take you so much time. If you
51:25simply create a readonly Postgress role
51:29and give your agents read only access to
51:32your production database, it's going to
51:34save you so much time. And anytime you
51:36build any feature, you check it. You
51:38give it a reality check. Does this
51:40feature even make sense? Does this even
51:42happen in production? Check the
51:44production database. Like, do people
51:45have this issue? Are they even using
51:47this feature? Should I be like improving
51:49this UI of this feature or is it like
51:50one of the least used features, right?
51:52absolutely essential, priceless. I wish
51:54I did this sooner. Maximizing of
51:56subscriptions. We already talked about
51:57that. Tracking agentic productivity. Uh
51:59actually, let me show you. We just
52:01released a new repo on this under
52:04Vectoral Labs. So, by the way, guys, I'm
52:05going to be posting less on this YouTube
52:07channel uh less often than I was before.
52:09But what I started doing is writing
52:10daily articles. So, if you want those,
52:12go to vectalabs.com
52:14and you can see that every day I'm
52:15releasing an article human written by
52:18myself. Um so, for example, yesterday I
52:21released it. Why aren't there more
52:22agents? Just trying to reason through
52:24it. You know, why are we running
52:25thousands of agents? The models are good
52:27enough, the harnesses are good enough.
52:29What is the issue? So anyways, this is
52:31the articles. But the second thing that
52:33Vector Labs does, which is you know my
52:35kind of research lab, basically
52:37experimental innovation lab is that we
52:40release open source projects. The newest
52:42one is agentic productivity which you
52:44can get find here. And this tracks your
52:47productivity. you know, how many commits
52:49are you doing, how many uh agent
52:51sessions are you running, how many user
52:52problems. I mean, each of these is kind
52:53of a bad metrics on its own, but when
52:55you combine these three metrics and you
52:57see a long-term trend across all
52:59different harnesses, then you can see
53:00like whether you're becoming a better
53:02agentic engineer in terms of your actual
53:04output. So, yeah, this is one of the
53:06repos that we released fully open
53:08source. And again, go to vectalabs.com.
53:10I'm going to link it below the video to
53:12access all of these uh projects, open
53:14source projects for free or just read my
53:16articles because as I mentioned, I'm
53:17going to be posting less YouTube videos,
53:20but I'm going I'm writing articles daily
53:22highly, you know, high signal articles
53:24on like the cutting edge of AI. So, if
53:26you're interested in AI coding, agentic
53:28engineering, or just staying at the
53:30frontier of AI, vectal.com
53:33is your place to go. With that being
53:35said, I think we can wrap it up here,
53:36guys. So, thank you so much for
53:37watching. Hopefully you enjoyed this
53:39kind of raw highle video of my edeting
53:41engineering setup. If you did, comment
53:43below and I can make a part two. Maybe
53:45in a month because this this changes all
53:47the time, okay? So, in a month it's
53:48probably going to be different. And
53:49again, if you don't have your own VPS
53:51and your own cloud agents, set up your
53:53own hosting VPS right now. Just go ahead
53:55and do it right now. It's the first link
53:56below the video. It's one of the most
53:58affordable options out there. It's super
54:00easy. I literally showed you how to do
54:01it with an agent in this video and
54:03you're going to be prepared for the
54:04future where all of us are going to be
54:06running hundreds and thousands of agents
54:08in the cloud and you're not going to be
54:10ecosystem locked in because you're going
54:12to have your own VPS. So go ahead and
54:14click to the first link below the video.
54:15With that being said, thank you guys for
54:17watching and I wish you a wonderful
54:18productive week. See you.

About this transcript

This page contains the full transcript of My Agentic Engineering Workflow (after 6,775 sessions) by David Ondrej, generated from the public captions YouTube serves with the video. The transcript has 11,345 words across 1,572 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.