YT YouTube2Text

YouTube2Text → Transcripts → Top #1 Opportunity for Senior Engineers: Agentic Engineering

Top #1 Opportunity for Senior Engineers: Agentic Engineering — Transcript

by IndyDevDan · 5,330 words · 789 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00What's up engineers? Handy Dev Dan here.
0:02This is going to be a raw video and it's
0:03more of a message to myself. I just got
0:06back from 2 weeks in Greece completely
0:08unplugged and I used the trip to compact
0:10my contacts around one question. What's
0:14the greatest opportunity available for
0:16senior engineers? I came home with one
0:18answer. The opportunity is the same it's
0:21been for over a year now, agentic
0:24engineering. The opportunity is the
0:25same, but the size of the opportunity
0:27continues to grow. So, we have to grow
0:30with it. Andrew Karpathy just called out
0:33agentic engineering directly at the
0:35Sequoia AI ascent. When Karpathy names
0:38something, the industry follows, which
0:41means the window where you're early is
0:42closing. By the end of 2026, agentic
0:45engineering will be the default. On this
0:48channel, we've been betting on agentic
0:50engineering since Claude code was
0:52released way back in March 2025. In
0:54fact, we've been earlier than that. I
0:57own the domain name agenticengineer.com
0:59where myself and thousands of engineers,
1:03some of the best engineers from top
1:05companies that you know, have
1:06accelerated past the industry by pushing
1:09the ceiling of agentic engineering, not
1:12the floor of vibe coding. It's not
1:14enough to be early, you have to keep
1:16showing up because the game is always
1:19changing. Here's the fact on the ground.
1:21Two engineers using the exact same agent
1:24with 200K tokens can get massively
1:27different results. In this video, I want
1:29to talk about the five pillars that make
1:32up the gap between low and high
1:34performing agentic engineers. These are
1:36the key ideas I'm focusing all my time
1:38and effort on right now, so I want to
1:40share them with you here. We're going to
1:42talk about agent harnesses, software
1:43factories, extensible software,
1:46always-on agents, and agentic access. If
1:50you're interested in using these ideas
1:51to get the next generation of results
1:53today, stick around. This is exactly
1:55what I'm thinking about every single day
1:57to win [music] by achieving the ceiling
2:00of what's possible. So, let's reset our
2:01context on the top one opportunity for
2:04senior engineers, agentic engineering,
2:06[music]
2:07the agent harness. Whoever controls the
2:09agent harness controls your results. Why
2:12do I say this? Let's break it down. Let
2:14me start with a crazy stat. I'm building
2:16one new custom agent harness every
2:18single day. It's catching up to the
2:20total number of skills and agents I
2:22create. Why is that? It's because I want
2:24more control over the agentic
2:26engineering experience, and owning your
2:28agent harness is exactly how you do
2:31that. Let's just be really, really raw
2:32about this. We're talking about cloud
2:34code, we're talking about Codex, we're
2:35talking about open code. These tools are
2:37fantastic. They were a great start,
2:39they're a terrible place to finish. Why
2:40is that? It's because the agent is
2:43everything. It's the agent that gives
2:45you the superpowers. The agent unlocks
2:47the agentic speed, right? You
2:48[clears throat] know this. Human speed
2:50is too slow. The agent lives in the
2:51agent harness, and therefore, whoever
2:53controls the agent harness controls your
2:56results. These are just the
2:57surface-level things that you'll control
3:00when you own your agent harness. As many
3:02of you know who tuned into the channel,
3:03I use the Pi coding agent. The Pi coding
3:05agent lets me build composable units of
3:09customization, okay? It allows me to
3:11build a stack of unique experiences that
3:14combine to build a net new agent
3:17harness. So, simple examples we've
3:18talked about on the channel, multi-agent
3:20teams. For those of you that are new to
3:22the channel, who are just tuning in
3:23right now, we don't just talk about the
3:24stuff, we actually build against it. So,
3:26I'm just open this up for you. UIA J
3:28team. This is a custom Pi agent harness
3:32where I've created a multi-team
3:33orchestration system. So, this is a form
3:35of multi-agent orchestration, right? So,
3:36I'll say ping your teams, then the
3:39orchestrator will ping all of its teams,
3:41set up brand UI generation, validation,
3:43and then the team leads also connect and
3:46coordinate with their workers. So,
3:48there's three tiers of agents here.
3:50They're operating in a chat room-like
3:52interface here, and as you can imagine,
3:54this is impossible with any
3:56out-of-the-box default agent harness
3:58that you're renting. What's another
4:00example of a customization harness? Let
4:01me show you my default out-of-the-box
4:03harness where I've stacked up and
4:05composed different slices. Here, we're
4:07pulling out skills, agents, commands
4:10from any location. Right now, it's just
4:12using that default since I'm in a temp
4:13directory, but you can also see here my
4:15agent communication system. I have
4:17agents operating on this network that I
4:19can directly prompt and talk to right
4:21from this agent, right? This is just one
4:23additional example of how you can
4:25customize and control your agent
4:27harness. Whoever owns your agent harness
4:29controls your results. I'm a huge fan of
4:31Cloud Code. I've been using it since the
4:33beginning, but this tool is the floor.
4:36It's not the ceiling. It's just the
4:37beginning of what's possible with tools
4:39like this. And so, I'm building a new
4:41agent harness every single day. It's
4:43crazy to say that, but it's true. I'm
4:44experimenting. I'm getting more control
4:46over the agentic engineering experience.
4:49Your tools directly shape what you
4:51believe is possible. With the Pi Agent
4:53Harness, I see no limits. You can see
4:56I've got two agents on this network
4:57already. I have a presentation Opus 4.7
5:00agent, and then I have a helper Gemini
5:023.5 flash agent testing out the
5:04capabilities of that new model that just
5:05dropped, and I'm using these two agents
5:06together on my agent network through my
5:09custom agent harness. This is part of
5:11the advantage you get when you own your
5:12agent harness, right? You want to
5:14control the experience of your agentic
5:17engineering. If you don't own the agent
5:18harness, you will always be limited by
5:20what another tool is telling you what
5:22you can and cannot do, right? So, you
5:23can create a sandbox tool. Every prompt
5:26you write, everything you do, your agent
5:28is operating in a sandbox environment
5:30just by default. Of course, you can
5:32recreate sub-agent delegation. You can
5:34add damage control like we've done on
5:36the channel in a very custom way. You
5:38can add model fallbacks. You can route
5:41to different models. There's just
5:42endless ways to use and build your own
5:44custom agent harness. And what does this
5:46give you in the end? It allows you to
5:48ultimately build with one tool, you can
5:50generate many versions. And again, I use
5:52the pie agent harness. Shout out to
5:54Mario for building such a great tool.
5:55And there are really kind of two classes
5:58of agent harnesses you can build. You
5:59can build engineering pattern focus
6:02agents like the pie-to-pie agent harness
6:04we talked about last week. I'll link
6:05that in the description. You can build
6:07agent chains. You can build the verifier
6:08harness we talked about a few weeks ago
6:10where you have an agent always checking
6:12over the work of another agent. But
6:13then, you can do something really really
6:15powerful and more important that many
6:16engineers are not tapping into. You can
6:18build domain specific agents that do one
6:20thing extraordinarily well. DevOps
6:22harness, a testing harness, a billing
6:25harness. You know, the specialization is
6:28truly endless. But if you don't own the
6:30agent harness, you cannot specialize
6:32your agentic engineering experience. One
6:35tool, many versions, specialization is
6:37the most This will always be true. If
6:40you can specialize the experience to
6:42outperform someone using an
6:43out-of-the-box agent for your product,
6:45for your service, for your application,
6:47you will win. Whoever controls the agent
6:49harness controls your results. So, you
6:51want to own that. So, what's another key
6:52pillar of agentic engineering I'm
6:54centering myself around to get that next
6:56set of results? It's of course the
6:58software factory. And the big idea here
7:00is you want to be building factories,
7:03not features. What does this give you?
7:05It gives you on-spec results every time.
7:07So, whatever you teach your agents to do
7:09is what you're going to be able to tap
7:11into over and over and over. Why are the
7:13best engineers using software factories
7:16and not just building the future
7:17themselves? When you are working on your
7:19software factory, you are building the
7:20system that builds the system. Instead
7:22of going direct to your new app feature,
7:25to your new landing page, to some new
7:27API you're building for a client, for
7:29some service, or setting up some
7:31accounting spreadsheet garbage, instead
7:33of doing any of that yourself, you are
7:34building the teams of agents, the
7:37systems of agents plus code that does it
7:40on your behalf. And so, you're moving
7:42the unit of engineering work into the
7:44system that builds the system. You're
7:45moving your focus into the factory. And
7:48the first place you see this is the plan
7:51prompt or the spec prompt. You give your
7:53agent a formula for how you create plans
7:57for how engineering work is done. Let's
7:59be super clear about this. What is a
8:00plan? It's a prompt scaled. That's all
8:02plan is. It's a more detailed prompt.
8:04But that's just the beginning. There's
8:05planning, there's plan reviewing,
8:07there's scouting, there's validating,
8:10there's the actual building of the
8:11thing, there's a testing, there's the
8:12reviewing. All of this is part of your
8:14software factory. And so, who cares? Why
8:17is the software factory so important?
8:18It's because your output per unit of
8:20time goes parabolic. This is what really
8:22matters. I've said this before, I'm
8:24going to keep saying it. There's an
8:25engineer right now that's prompting back
8:26and forth in their terminal. Maybe
8:28they're running two, five, 10 agents.
8:30The number doesn't really matter because
8:32the other engineer, you and I, we're
8:34going to be building factories. And when
8:37you build the factory, you can reproduce
8:39that focused result over and over and
8:42over. You write one prompt and your
8:44factory of agents and code work together
8:47to produce a massive result, to produce
8:49an entire feature, to produce an
8:51outcome. This frees you from being stuck
8:54inside the terminal, which ties directly
8:56into another key pillar, AFK agents.
8:58We'll get to that in a moment. But if
9:00this idea makes sense, drop a like and
9:02subscribe because this means that you
9:04understand the true leverage available
9:06for agentic engineers. You have to
9:09invest the time in to your software
9:11factory. And this is also known as the
9:12dark factory in tactical agentic coding.
9:14We call these ADWs, AI developer
9:17workflows. And the whole idea is to
9:19combine agents plus code to outperform
9:21either alone to get repeat results in
9:24your system, not anyone else's, in your
9:26system. You teach your agents to build
9:28just like you when you build your
9:29factories. You template your
9:31engineering. It allows you to produce
9:32that high-quality car over and over and
9:35over. The a
9:36result over and over and over. Just to
9:39kind of put the stamp on this, it's your
9:41result. You're creating a stamp, a
9:43system of reproducible results that's on
9:46spec every time. Your test, your
9:48validation, when you build a factory, it
9:50always runs. Always getting that
9:52verification, you're releasing just like
9:53you would. Maybe you have a staging
9:55environment here that you can preview to
9:57validate. Maybe you have a whole team of
9:59agents that runs and fixes all
10:01regressions. It's this system, right?
10:03It's that factory of results that you're
10:06creating that you can tap into when you
10:08build the software factory. So, this is
10:11an idea I'm spending a lot of time
10:13thinking about. You don't build the
10:15feature. This is a mindset shift. I'll
10:17be honest with you, this is hard to do.
10:18You have to shift yourself from
10:19thinking, I'm the engineer that builds
10:21the feature. No, you're the engineer
10:23that builds a system of AI plus code
10:27that operates on your behalf. You're
10:28building the software factory. You are
10:30building the system that builds the
10:32system. This is the key thesis inside of
10:34Tactical AI Coding. And on the channel,
10:36we talk about this all the time because
10:37it's so important. And I want to be like
10:39really, really blunt with you here. The
10:41industry, right? All the top 500
10:43companies, everyone is building a
10:45software factory. Soon, this will be a
10:47requirement. This is how you get massive
10:49leverage out of your tokens. By the end
10:51of the year, all of your products that
10:53are in production, you should have
10:54software factories. You should be able
10:55to write a prompt and then see a result
10:57near production. And if you're doing
10:59really, really well and you're really
11:01owning this, you might be getting close
11:02to ZTE, zero touch engineering where you
11:05go from prompt directly to production.
11:07That's super, super advanced. We're not
11:09going to focus on that here, but that's
11:10where this all goes. All right? ZTE,
11:12zero touch engineering. But none of that
11:14matters if you don't have the factory
11:16that can build the features for you on
11:17spec every single time. This is not a
11:19black or white thing. This is a gray
11:21thing. It takes a lot of practice, a lot
11:22of agentic engineering. This is a new
11:24skill, just like full stack, just like
11:26back end, just like front end, just like
11:28DevOps. This is a software engineering
11:29skill that takes practice. There's a
11:32practice in building agents that can
11:34operate for you on your behalf. But the
11:36end result is you get a software
11:39factory. Let's move on to the next big
11:40idea. Extensible software. I would say
11:42if there are two ideas that I've
11:44personally missed, it's controlling your
11:47agent harness and it's extensible
11:49software. Being fully honest transparent
11:51here, these are the only two ideas I
11:52missed in Tactical Agent to Coding,
11:54which is why I've added and extended the
11:56course into Agent Horizon. And this why
11:58I'm talking about um the Pi Agent
12:00Harness so much on the channel right
12:02now. But it's not just about the
12:04harness, right? This is about software
12:06in general. You want to be writing
12:07extensible software. It's very very
12:09clear why. It's probably not a surprise
12:12to you as a senior engineer. It's really
12:14about the number of changes that's
12:15happening in the industry. We have model
12:17changes, we have new prompts, new tools,
12:19new technology. We have change happening
12:22all the time. And the best way adapt to
12:25change is to build it into your
12:27software. You want to be building
12:28pluggable software. Ease to extend, easy
12:31to change, easy to tweak. It needs to be
12:33adaptable. Because as you know, models
12:36will change, tools will change. These
12:37things are are moving at light speed.
12:39The pace will only continue. Software is
12:42going parabolic right now. GitHub is
12:43crashing all the time. Critical services
12:46changing all the time. Because we'll be
12:48operating at the agentic speed. And so
12:50if your software's complex, if your
12:52software's slow, if it's brittle, if you
12:54are vibe coding trash, you have not seen
12:56or read, if you're generating more slop,
12:58creating more tech debt in your
13:00technology, you will suffer.
13:02>> [laughter]
13:04>> Extensible software is something I'm
13:05thinking about a lot right now. And
13:07again, there's two kind of classes for
13:09this. There's engineering work and
13:10there's product work. There's enhancing
13:12your development and a great example of
13:14this is your agent harness. And then
13:16there is production work. Whether you're
13:18deploying agents into production for
13:20agentic products or if you have
13:22traditional software where, you know, AI
13:24isn't involved, it doesn't really
13:26matter. You need adaptation. Agent
13:28Harness is a great example of this,
13:29right? Constantly, there are different
13:31tools, there are new prompts, there's
13:33new agents you want to be testing,
13:34different system prompts. You want to be
13:35able to control the model that you're
13:37using all the time. This is why the Pi
13:39Agent Harness is so important. It's so
13:41swappable, it's so dynamic. You can
13:43change this thing on the fly. When
13:44you're thinking about extending your
13:46lead as an agentic engineer, I highly
13:48highly recommend you think about how you
13:49can make all of your software more
13:51extensible. Plugability is the key.
13:54Extensibility is the key. Composability
13:57is the key. If your software has a
13:59million trillion rules and operates in a
14:01very specific line of cascading if
14:04statements, the next year is going to be
14:06really really hard for you because your
14:07agents will be really slow, they'll be
14:09making a lot of mistakes, and you have
14:10to teach them in your software how to
14:12navigate all that. If you have
14:13extensible software, adaptable software,
14:15if you are adding and not modifying, you
14:18will have a massive advantage. This is
14:20one of my favorite software principles,
14:21open to extension, closed to
14:23modification. All right, so I'm thinking
14:24about that a lot as I use agents every
14:26single day, as I'm building both a new
14:28agentic engineering workflows for myself
14:32to improve and increase my speed, but
14:34also on the product side, thinking about
14:36how to build great products that survive
14:38and thrive in a world where change is
14:41the default mode. Build extensible
14:43software, all right? Makes sense. And if
14:45you're a senior engineer, you know this.
14:46These are ideas I'm re-emphasizing as
14:48I'm resetting my context for this next
14:51phase of work. Always on agents. When we
14:53talk about the ceiling of agentic
14:54engineering, this is one of the most
14:56important ideas. It's not enough to just
14:58have an agent that's always running.
15:00Anyone can set up a cron job, spin up an
15:01agent, run a while loop on an agent, and
15:04just have it always be burning tokens.
15:06The key here is useful tokens. In
15:09Silicon Valley, in a lot of these tech
15:10companies, there's a term called token
15:12maxing. You've probably heard of it.
15:14Token maxing sits at the lowest level of
15:17our economic funnel of tokenomics. Okay,
15:19so there's three levels here. First, you
15:22need to use more tokens. Then, you need
15:23to make your tokens useful. And then,
15:25you capture the revenue generated by the
15:28value you created. Okay, so these are
15:30the three levels. Now, if you're token
15:31maxing, you're very likely down here.
15:34You're just using more tokens. This is
15:36great. It's a great place to start,
15:37terrible place to finish. Once you're
15:38spending tokens in your agentic systems,
15:40the next step is to make them valuable.
15:43And I think this is what, you know,
15:44where a lot of companies, a lot of
15:46engineering teams, a lot of adopters of
15:48agents are really sitting right now. A
15:50lot of tokens getting generated, not a
15:52lot of value getting generated. This is
15:54the trick of engineering. Only after you
15:56do that are you able to move to that
15:57highest level. Now, why am I saying
15:59this? Why is this important? It's
16:00because once you generate valuable
16:02tokens, and once you can capture that
16:04arbitrage, the next step, the only clear
16:06step is to turn them on all the time.
16:09This is the ceiling of agentic
16:10engineering. Agents that are on for you,
16:12operating, working while you sleep. But,
16:15in order to get there, you must
16:17understand the token arbitrage. This is
16:19what everyone with their head on their
16:20shoulders is trying to do. If you're a
16:22senior engineer looking at the best way
16:24to spend your time, it's following this
16:26equation. We're all purchasing tokens
16:28right now. If you are at level two,
16:30you've made them useful. Then, you roll
16:31them into your products or your
16:33engineering experience. Once you've
16:35generated value for your users or for
16:38yourself, the next thing to do is to
16:39capture the value, right? You need to
16:41capture the actual revenue generated by
16:44your tokens. And so, it's these three
16:46levels that matter. First, you use more
16:47tokens. Then, you make your tokens
16:49valuable. And then, from your tokens
16:51that are valuable, you need to turn that
16:52into real revenue. That's the arbitrage
16:55that's happening right now. This is what
16:57everyone is and should be focused on
16:59right now. You buy the tokens at a
17:00certain price. You make them useful. You
17:02roll them into your products and your
17:03engineering experiences. And then, you
17:05need to capture the actual revenue
17:07generated from them. This is the token
17:08arbitrage. This is what matters. This is
17:10tokenomics. And then, the whole key is,
17:13once you can successfully do this, you
17:15turn it on. You have always-on agents,
17:17AFK agents. When you do this right, you
17:20then have a new KPI, right? Your rising
17:22API bill is a new productivity KPI, but
17:25only after you get out of level one and
17:26level two. Because then this turns into
17:28a classic business problem. How many
17:30tokens can you spend? What are they
17:31worth to the business? And then you
17:33capture that difference. So if you pay a
17:35dollar for a token and you can, through
17:37your business, generate a dollar and 10
17:40cents worth of value for that token and
17:42then you capture that 10 cents, right?
17:45You have an infinite cash generating
17:46glitch. Also just known as a business.
17:49And then all you need to do after that,
17:51just like when you have a great ad
17:52campaign that's running, you just scale
17:54it to the moon, right? You spend as many
17:56tokens as possible. This is literally
17:58what you see in the big AI labs. With
18:00more tokens that they can sell, they
18:02make more money. But contrary to popular
18:04belief, I believe this goes all the way
18:06through the stack, all the way down to
18:08companies, all the way down to
18:09individual developers like you and I. If
18:11you can buy a token for a dollar, run it
18:13through your business process, and sell
18:14it for $2, that's a 2x right there. And
18:17then all you need to do is smash tokens
18:19and scale it up. So it looks something
18:21like this. Your API cost should go up
18:22and the value you're generating should
18:24go up directly with it, ideally greater
18:27than the cost. It doesn't need to be a
18:28linear scale, it most likely will be,
18:30but this is a simple idea. And once you
18:32do this, the whole idea is to scale it,
18:34right? Turn that system, turn that
18:36agent, turn that product on all the
18:38time. And so I'm thinking about this.
18:40And the trick here is they have to be
18:42useful tokens. Token maxing, once again,
18:44it's the floor. I can spend millions of
18:46tokens right now. I think you would be
18:48very, very surprised to see my token
18:50usage. It's actually quite low. My token
18:52growth is a very, very smooth curve
18:54because once I find that key value, that
18:57arbitrage, once I unlock value
19:00generation and turn that into revenue
19:02capture for whatever business, products,
19:04clients I'm working with, that's when
19:05you scale it to the moon. But it's only
19:07at level three. If you're doing it at
19:08level one or level two, you're You're
19:10token maxing. Once you do that, you turn
19:12that agent always on. You turn that
19:13product, that system always on, that
19:15agent harness, turn it on. But, always
19:17think through that first. And this is
19:19the key difference between that
19:20high-performing engineer using the exact
19:23same agent and the exact same number of
19:25tokens. If you're generating more value
19:27and you're capturing the value from the
19:29tokens you generate, you will be ahead
19:30of the pack. You will be capturing more
19:32revenue and therefore delivering more
19:34value. And only then do you turn it into
19:37an always-on agent. There are a million
19:39agent cron jobs running. 90% of them are
19:41dead useless and just burning cash,
19:43burning tokens. Agentic access. So, the
19:45headliner here is simple. You likely
19:47already understand this idea, but it's
19:49very important to emphasize because I
19:52know for a fact that you, a senior
19:54engineer, you probably haven't given
19:56your agent API access to all the things
19:58you really could to maximize your
20:01agentic speed. API access is a
20:03requirement of agentic speed. Let's just
20:04quickly go through this. The ideas here
20:06are simple. You already understand this,
20:07but it's really important to to
20:08re-emphasize. As I'm refocusing my
20:11personal context for this next phase of
20:13work, I'm thinking about always moving
20:15at the agentic speed. If there's
20:17something an agent can do, uh you have
20:19to really deeply ask yourself why it
20:21isn't it. And a lot of the times, it's
20:23just because you haven't given your
20:24agent the API access it needs. But, this
20:27is how you get the agentic speed, okay?
20:29And this is the only way to operate now.
20:31CLI tools, REST, webhooks, RPC clients,
20:34you know the deal. Agents only command
20:35what they can programmatically reach.
20:37And in order for them to act and operate
20:39like you, you need to give them the
20:40tools that you have access to. There are
20:42limits here. You don't want to be giving
20:44production access so that they can nuke
20:46your production databases, your volumes,
20:48your resources. As talked about in a
20:49recent video, you want to lock down that
20:51bash tool so that your agents never do
20:53anything catastrophic. But, the idea is
20:55very simple. You want to avoid the token
20:57tax. So, what is a token tax? It's
20:58basically anything your agent does that
21:01is unnecessary strictly because you
21:03haven't given it direct API access. You
21:05want your agents using tokens only when
21:08they need to to do the thing that is the
21:09most useful and valuable to you, your
21:12company, your product, your users. Don't
21:14pay a token tax. And this is that simple
21:16mindset shift that you already know. I
21:17know that you know this. API access is
21:20requirement of moving at the agentic
21:22speed. Agents they're 10, hundreds,
21:24thousands of times faster at operating
21:26digital information than you or I. This
21:28is also why we don't work on the
21:29feature, we work on the agents that
21:31build the feature, right? We build the
21:32system that builds the system. It's the
21:33same thing. We want to be moving at the
21:35agentic speed, the speed of agents. But
21:37you can't do that if they don't have API
21:40access. So once again, I am refocusing
21:42my context around this. There are things
21:44you and I are doing right now we don't
21:46have to be, but we are because we just
21:48haven't invested the time to give our
21:50agents the tools, the access, the
21:52resources that they need. So make sure
21:54you do that exactly, okay? So these are
21:56the five pillars that are going to
21:57compound your advantage toward the one
22:00opportunity that matters the most right
22:01now, okay? It's agentic engineering.
22:03Capacity set it, the entire industry is
22:05catching up with this. You're going to
22:06hear these kind of similar terms come
22:09around all the time, right? Dark
22:10factories, software factory, AI
22:12developer workflows, always-on agents,
22:14AFK agents, agentic access,
22:17extensibility, agent harnessing, right?
22:19Harness engineering. These are all ideas
22:21that are going to come up over and over
22:22and over in different forms and shapes.
22:24As I've been saying on the channel,
22:26success is often about doing a few very
22:29simple things over and over and over and
22:31over and over. And and you've seen this,
22:33you know this. That the name of the
22:35game, the one opportunity is the thing
22:37you need to focus on, and I'm saying
22:38you, but really as I mentioned, I
22:40created this presentation, a lot of the
22:42work I do is like a note to myself, it's
22:44a reminder to myself. You want to be
22:45agentic engineering, you're going to be
22:47building the system that build the
22:49system. You want to be focusing on
22:50pillars that compound, actions that
22:52compound, so that you can really tap
22:54into the ceiling of agentic engineering.
22:56Stack your advantage, compound the
22:58opportunity, raise the ceiling of your
23:00agentic engineering. Vibe coding is the
23:02lowest-hanging fruit. Do not sit in the
23:05terminal prompting out your features.
23:07Build the software factory, own the
23:09agent harness, use extensible products
23:11like the pie agent harness, but also
23:13make your products extensible by nature.
23:16Build it into the fabric of the work
23:18you're doing. Learn to arbitrage your
23:20tokens. You're spending a bunch of
23:21tokens that do not generate value is
23:23meaningless. You need to spend more
23:25tokens, you need to figure out which
23:26ones are actually valuable, and then you
23:28need to capture the value usually via
23:30revenue from your tokens. Then, once you
23:33do that, you turn that thing on so it
23:35runs 24/7. Always-on agents, AFK agents,
23:39get that maximum value you can out of
23:41that system. And then, you know, more of
23:42a reminder than anything, if there's
23:44something you're doing that your agents
23:46can do, why aren't you?
23:48Right? Give your agents access. Make
23:51sure always focused on agentic access,
23:54agent-first systems, agent-first
23:58products, agent-first workflows. Expose
24:01your CLIs and APIs everywhere. Code
24:04bases, products, devices, tools, agents
24:06can only command what they can reach. If
24:08you're not doing this, but you're still
24:09using agents, you're paying a token tax.
24:12You're burning up tokens because you
24:14didn't put the upfront investment into
24:16the system that you need to. Again, this
24:18is more of a reminder to myself. A lot
24:20of the things I do on the channel,
24:21they're for me as much as they are for
24:24you. But, this is what I'm working on
24:25every day as an engineer right now. The
24:27gap between, you know, the top 2% of
24:29engineers and everyone else is widening
24:31every single week. Why is that? It's
24:33because they're compounding their
24:35advantage with this one opportunity.
24:37It's very rare that you're presented
24:39with an opportunity to compound an
24:41advantage. Don't miss this. Move slow
24:43now to move fast later. Invest in your
24:45agentic layer. Don't sit in the terminal
24:47vibe coding. Agentic engineer your work.
24:49Build the system that builds the system.
24:52I have a bunch of links that I'll share
24:54in the description for you. Previous
24:55videos where we talk about the pie to
24:57pie customize agent harness of course
24:59share tactical agentic coding for anyone
25:02that's interested in understanding what
25:04is possible right now with agentic
25:06engineering and what will be possible
25:08once you put the reps in a big shout out
25:10to Mario and everyone working on the pie
25:13agent harness this tool has been
25:14incredible to work with as I mentioned
25:16I'm building a new agent harness every
25:18single day now which is really absurd
25:20you can imagine
25:22I have a software factory that lets me
25:24do this so bunch of links in the
25:26description for you check all them out
25:28bunch of free stuff some paid stuff if
25:29you're interested in that extra push
25:31these are the five big ideas I'm
25:33focusing on I'm re-centering myself on
25:35I'm emphasizing in all of my work and so
25:37I recommend you do the same comment down
25:39below if you think I missed anything
25:40you'll notice here I didn't mention
25:42models a single time that's because
25:44models matter less and less they matter
25:46for bleeding edge work but for a lot of
25:49the work 80 90% of the work we're doing
25:51on a daily basis what matters is the
25:53systems you place around your agents we
25:55are talking about agentic engineering
25:57which to be clear is the process of
25:59engineering with intelligence that can
26:01operate on your behalf so five big ideas
26:03for you I've got more resources linked
26:05for you in the description shout out pie
26:07shout out Mario shout out Andrew
26:09Karpathy shout out Claude code
26:12>> [laughter]
26:12>> and thank you for watching to the end
26:15here you know where to find me every
26:16single week stay focused and keep
26:19building

About this transcript

This page contains the full transcript of Top #1 Opportunity for Senior Engineers: Agentic Engineering by IndyDevDan, generated from the public captions YouTube serves with the video. The transcript has 5,330 words across 789 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 choose a caption track available on the video.

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, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.