0:02 So there are essentially infinite AI
0:03 jobs right now. Not growing demand, not
0:05 a hot sector. None of that is true. It
0:07 is functionally infinite. And that's
0:09 true for businesses employing 10 or 20
0:10 people as much as it's true for
0:11 businesses employing hundreds of
0:14 thousands. There is no functional upper
0:17 limit to what employers would love to
0:19 have as far as AI talent and they cannot
0:22 find them. After hundreds of interviews
0:24 on particular roles, I am hearing from
0:26 employers, we can't fill the role. And I
0:29 hear you saying, "Nate, that must be a
0:31 lie. I have applied for hundreds of AI
0:34 positions. I am good at AI. It is not
0:36 working." I get it. We have a K-shaped
0:38 job market right now. And there is a
0:41 split in what employers want that is
0:43 confusing this issue because many
0:46 employers who don't fully know AI are
0:48 taking advantage of this situation by
0:50 basically putting resumes out as
0:52 learning tools and then they use the
0:56 interviews to learn from the talent what
0:58 they need which is really terrible right
1:00 like that's not the way you should be
1:02 interacting with talent. It doesn't
1:04 bring the best people to you and it
1:06 leaves a bad taste in everyone's mouth.
1:09 There are also plenty of people who are
1:10 looking for roles who are either
1:13 overstating their capabilities or who
1:16 don't have the actual skill sets needed
1:17 to thrive in AI. I'm not talking about
1:19 being able to chat with the AI, right?
1:21 I'm going to give you in this video
1:24 seven specific skill sets that I have
1:26 pulled from looking at hundreds of
1:29 actual AI job postings and then looking
1:31 underneath at what the subsklls building
1:32 those skills are. And I've gone farther
1:34 than that, right? I'm not just going to
1:35 talk about the skill sets. I'm going to
1:37 talk about how you develop them. I'm
1:40 going to put up a guide on the Substack
1:42 that helps you to actually get to those
1:43 skills. And I've gone even farther than
1:45 that. I'm putting together a hiring
1:46 board because I think it's really really
1:49 confusing right now to have AI talent
1:51 and AI hiring managers mixed in with
1:53 everybody else because then you're like
1:55 looking among all of these PM jobs.
1:57 Which are the AIPM jobs? Which are the
1:59 whitewashed AIPM jobs where it's just
2:01 like we say AI but we don't mean it. And
2:02 I'm trying to fix that because I think
2:04 it's time to bring some simplicity to
2:05 this. Hiring doesn't have to be as hard
2:08 as it is. Okay, before we get into all
2:09 of that, let's dive in to what's
2:11 actually going on in the job market.
2:13 Fundamentally, the AI labor market is
2:15 actually two markets moving in opposite
2:17 directions. When I talk about K-shaped,
2:19 I mean it. Market one is the traditional
2:20 knowledge work roles, right? The things
2:22 that we've all learned since the 2010s,
2:24 generalist product managers, standard
2:26 software engineers, conventional
2:27 business analysts. And there's no other
2:29 way to say it. I'm sure you're not
2:32 surprised that job opening count is flat
2:34 or falling. It is not growing because
2:36 most of the interest, most of the
2:37 investment where businesses are
2:39 investing to grow, it's on the AI side.
2:41 And that's the other side of the market.
2:43 It's roles that design, build, operate,
2:45 and manage AI systems. And that is
2:48 growing fast. In fact, I've been kicking
2:49 around tech for multiple decades, and I
2:51 have never seen it this hot for this
2:53 kind of a job family. The ratio of AI
2:57 jobs to AI talent right now is 3.2 to1.
3:01 In other words, there are three plus AI
3:04 jobs for every single qualified
3:06 candidate right now. They can command
3:07 their price and they do. If you want
3:09 specific numbers, this is from a
3:11 manpower group survey which found 1.6
3:13 million jobs, which I think is low, and
3:15 only about half a million qualified
3:17 applicants, which I think is pretty
3:19 fair. And that's leading to a very long
3:22 time to fill the role. 142 days to fill
3:24 an AI role, which is almost half a year.
3:26 And so the people who tell me I'm lying
3:28 when I say that this market exists, I
3:30 get it. If you're not in that half a
3:31 million category, it does feel
3:33 impossible because the entire rest of
3:35 the job market is condensing into
3:37 commodity. But if you're on the other
3:39 side, you get it. This is a world where
3:41 you can write your own ticket because
3:42 people are desperate for these skills.
3:44 So without further ado, I don't want to
3:46 belabor this. Let's get into what these
3:48 skills are. I want this to be the most
3:50 useful video on these skills because I
3:52 went and I looked at all of these AI
3:53 courses before I made this video cuz I
3:56 was like surely someone out there has
3:59 made a video that is based empirically
4:01 on the AI job postings, backward
4:03 analyzes them, decomposes them into
4:05 subsklls and gets very specific about
4:07 what employers are hiring for. That is a
4:09 learnable skill. And by the way, this is
4:12 easier than other information tech
4:13 revolutions. If you think this is hard,
4:15 when you were getting a personal
4:18 computer in the 1980s to learn how to
4:21 code, you had to fork over like 15 or
4:23 16,000 in today's dollars to do that. It
4:25 was ridiculously expensive. It was
4:27 heavily gated by your ability to afford
4:30 stuff. Now, it's much much easier.
4:31 Almost anyone has access to an AI
4:33 subscription if they want. AI can
4:35 actually help you learn. We can do this.
4:36 And we're going to start with the most
4:38 fundamental shift of all. People
4:39 sometimes call this prompting. I've
4:41 talked a lot about prompting. I want to
4:44 use the term that I am seeing more and
4:45 more in job postings and that is
4:48 specification precision or clarity of
4:51 intent. You have to learn to talk
4:54 English to a machine in a way a machine
4:56 takes literally. We are used to working
4:57 with humans that read between the lines.
4:59 We're used to working with humans that
5:00 can infer from our intent pretty
5:02 reliably. One of the reasons we know
5:05 that general intelligence is not really
5:07 here yet is that agents don't do a good
5:09 job of that. Agents need us to be
5:11 specific. An agent is going to take
5:13 whatever specification you give it and
5:15 go and build something. And if you're
5:17 not clear about what that is, the agent
5:18 is going to try its best to fill in the
5:20 blanks, but that won't reliably
5:23 reproduce your intent. Agents are bad at
5:25 filling in the blanks. And yes, I'm
5:27 going to give you a specific example.
5:28 Let's say you're trying to improve
5:30 customer support, but you're not giving
5:31 something to a principal engineer where
5:33 you say, "Hey, come up with a solution
5:34 on customer support. You've read the
5:36 tickets." We're not going to be that
5:38 vague. Instead, we're going to be clear
5:40 about what we care about in the prompt
5:41 to the agent. This is the difference
5:44 that job posters are looking for. You
5:46 need to be able to say to the agent, I
5:49 want you to build an agent that handles
5:51 tier one tickets. I want you to be able
5:53 to handle password resets. I want you to
5:55 be able to handle order status
5:56 inquiries. I want you to be able to
5:58 handle return initiations. I want you to
6:00 know when to escalate to a human based
6:03 on customer sentiment. And I want to
6:05 define customer sentiment in such a way
6:07 for you here in these docs that you know
6:08 how to measure it and score against it
6:10 and escalate appropriately. I want you
6:12 to log every escalation with a reason
6:14 code. You have the same intent here, but
6:17 you notice how specific that is. That is
6:20 what the bar is for prompting in 2026.
6:22 You have to be able to be that clear in
6:24 your intent. Now, if you're a technical
6:26 writer, if you're a lawyer, if you're a
6:28 QA engineer, a lot of this is going to
6:30 feel super familiar because you've done
6:32 this kind of technical writing before.
6:33 The gap is shorter than you think. For
6:35 many of us who are not used to writing
6:37 this specifically, it is a new skill,
6:39 but it's absolutely learnable. All it
6:42 takes is understanding in detail what
6:43 you intend to put together. And I'm
6:44 putting these in a specific order
6:46 because this is actually the order you
6:47 intuitively learn them in. I'm I'm
6:49 putting them in a sequence that makes
6:50 sense for you. Once you specify what you
6:52 want precisely, you immediately run into
6:54 the next problem, which is did you get
6:56 it right? Did you get what you wanted?
6:58 We call that evaluation and quality
7:00 judgment. And it's the single most
7:02 frequently cited skill across all of the
7:04 job postings I've come across. I'm not
7:05 sure employers all get it. And I'm going
7:07 to define it really clearly here. This
7:08 is something, by the way, that is in
7:10 engineering job postings and ops job
7:12 postings and PM job postings. People
7:14 talk about having an agentic evaluation
7:15 mindset, whatever that means. And they
7:16 want you to be able to do automated
7:19 evals and simulation runs, etc., etc.
7:21 Upwork has job postings that demand
7:23 evaluation harnesses for functional task
7:25 and longitudinal metrics, right? They're
7:26 talking about building ways to test
7:29 whether AI does a good job. Every single
7:30 posting will use slightly different
7:32 versions of this, but really it comes
7:33 down to being able to build systems that
7:36 encode evaluation and quality judgment.
7:38 And this is what all of that taste
7:40 discourse is all about. It's just
7:42 dressed up in skill language. And
7:44 really, I get why people are pushing
7:45 back on taste because when you talk
7:47 about this as taste, it just feels vague
7:49 and unactionable and it just strokes the
7:50 ego. But really, what we're talking
7:52 about is error detection with a degree
7:54 of fluency. AI has really different
7:56 failure modes from human failure modes.
7:58 AI is often confidently wrong. It's
8:00 fluently wrong. Whereas humans, when
8:02 we're wrong, we tend to stumble. There's
8:03 a lot of tells we have that we're used
8:05 to hearing and seeing in other people
8:07 that don't show up with AI. And so, if
8:09 we're not used to working with AI, we
8:11 may incorrectly see the confident
8:13 response AI has and assume that that's
8:15 true and right and correct. And I see
8:17 this a lot, by the way. If you think
8:19 this doesn't happen, I have seen it
8:21 happen in real presentations where
8:23 people will say, "Well, the AI presented
8:25 it and it looked correct to me and look,
8:26 it has all the right headers and this
8:27 and that." And I'm like, "Yeah, but use
8:29 some critical thinking here. This isn't
8:30 actually correct. I don't care how
8:32 confident the AI was." The skill here is
8:35 resisting the temptation to read fluency
8:38 by the AI as competence or correctness.
8:41 It's just not. Another subskll here is
8:43 what I call edge case detection. You can
8:45 show that you understand a subject
8:48 deeply when you are able to look at the
8:50 response from the AI and say you know
8:52 this is correct at core but the edge
8:55 cases are wrong. I think anthropics
8:57 engineering blog actually does a really
8:59 good job of explaining how this taste is
9:01 actually a learnable skill. What they
9:05 say is a good eval task is written when
9:06 more than one engineer looks at that
9:09 eval task and would come to the same
9:10 conclusion on a past fail basis. In
9:12 other words, excellent evaluations are
9:14 something we can all agree on and we can
9:16 all learn to write. If you're an editor,
9:18 if you're an auditor, these are the
9:20 kinds of skills you're using all the
9:22 time. You're just applying them in new
9:24 ways. If you're not, this is the gold
9:26 standard skill. Skill number two here,
9:27 this is the one that's mentioned the
9:29 most, and we all are going to have to
9:30 get good at it, whether or not we have
9:32 engineering in our title. And really,
9:34 the the best and simplest way to get
9:35 good at this is to start reviewing AI
9:37 output as if it has your name on it.
9:39 Care about it. Insist that it be
9:41 correct. insist that it be right and
9:43 then as you start to build agentic
9:44 systems which is a learnable skill we'll
9:46 get into. You should be able to build
9:48 them in such a way that you can sniff
9:49 out the quality at the end. And speaking
9:53 of multi-agent systems, let's talk about
9:55 what skill is involved when we do these
9:57 complicated multi-agent systems because
9:58 people sometimes look at that like
10:00 that's a chasm they can't cross. Like
10:02 I've had people who say I can use chat
10:04 GPT, I can even use cloud code but when
10:06 you say multi-ent like go white at the
10:09 roots, right? Like it's not easy. It is
10:11 easier than you think. Fundamentally,
10:13 the skill of working with multiple
10:16 agents is the skill of decomposing tasks
10:18 and delegating. It's a managerial skill
10:20 and you can learn it. You just need to
10:22 be able to break apart work into
10:24 manageable segments. That is part of how
10:27 you understand what works. And then you
10:28 can pair that with some of these other
10:30 skills you're learning like specifying,
10:31 like writing evals to actually get what
10:33 you want done. Now, if you think this
10:35 sounds like regular project management,
10:37 it's not. Agents work so differently
10:40 from people. Agents need very defined
10:42 guard rails and infrastructure to work
10:45 correctly. You can give your team of six
10:47 a set of assignments that are decomposed
10:49 rather vaguely in human terms and they
10:51 will still figure it out. We're sort of
10:53 generally flexible as workers. You
10:55 cannot do that with agents. You have to
10:57 very clearly specify the goal, very
10:59 clearly specify your initial intent,
11:02 very clearly define how you want a
11:04 multi- aent system to run. And there's
11:06 not that many ways to do it. The current
11:08 best practice is to have a planner agent
11:10 that keeps a record of tasks and that
11:12 can work with a variety of sub aents to
11:13 get those tasks done. Now, if you've
11:15 ever broken large projects into work
11:17 streams, take comfort. That is a skill
11:19 that transfers because you're really
11:21 thinking through what are the logical
11:23 delineations. What are the chunks in
11:25 this workstream and how do we hand them
11:27 off? That is something that you can
11:30 learn to work with AI to do and AI can
11:32 help you on when you start to build
11:33 bigger projects. One of the most
11:35 interesting subsets of this skill right
11:38 now is the ability to know is a given
11:41 project correctly scoped for the agentic
11:43 harness I have. And I have videos up
11:45 where I talk about this. The idea that
11:47 you need to size your work for the
11:49 agentic harness you have. If you have a
11:51 singlethreaded agent harness that's
11:52 designed basically to be a little
11:54 engineer in the computer that works for
11:56 you, you have to size your tasks and
11:59 decompose them to fit that engineer in a
12:00 box. If you have a multi- aent system
12:01 and you have a planner agent that
12:03 operates over a long period of time and
12:05 it has sub agents, you have the
12:07 flexibility to define a larger task, you
12:09 still need to be clear enough about the
12:11 subtasks and their logical relationship
12:13 that the planner can make good choices.
12:15 And so this is something where I say
12:17 this out loud and you might think, "Oh,
12:19 this isn't that hard." I promise you
12:21 it's hard. I promise you it's technical.
12:24 I promise you it's learnable and people
12:25 will pay for this. This is something
12:27 that people are in desperate demand of
12:29 all over the world. Now, you might
12:31 think, well, this sounds really hard and
12:33 these agent systems probably fail. That
12:35 brings me to the next skill. It's called
12:37 failure pattern recognition. And it's
12:39 absolutely critical. It shows up in lots
12:40 of these job postings because when
12:42 employers put these skills together, you
12:44 know what they recognize? Wow, it's not
12:45 so easy as I thought, right? We have
12:47 lots of ways that agents fail. I need
12:50 someone who can diagnose this at root,
12:51 fix it, and get me back to being
12:53 productive. And yes, if you're wondering
12:55 if I'm going to get into detail here,
12:57 you got that right. Because to be honest
12:59 with you, failure pattern recognition is
13:00 not widely understood. And people tend
13:02 to say, "Well, oh, what's failure?" I'm
13:03 going to tell you, I dug into the
13:05 research. I also have seen these
13:07 failures. These are the six failure
13:09 types that pop up. Context degradation
13:11 is one, right? Quality is going to drop
13:13 as your session gets long because you're
13:15 polluting the context window. Another
13:18 one is specification drift. Over a long
13:20 task, the agent effectively forgets the
13:22 specification unless you construct your
13:25 agent harness correctly and the agent is
13:27 forcibly reminded of the specification.
13:30 A lot of what you see in the Ralph loop
13:32 on Claude that went viral is forcible
13:34 reminder of specification. Sycopantic
13:36 confirmation is another one. That's
13:37 where the agent actually confirms
13:39 incorrect data and then comes back and
13:41 builds an entire incorrect system around
13:43 that data. You have got to watch the
13:44 data you put into these agents. They
13:46 will take it seriously. they will
13:47 confirm against it. They will
13:49 sickyantically agree with it. And if you
13:51 are feeding them bad company data,
13:53 you're going to get bad systems. Tool
13:55 selection errors another one. Tool
13:57 selection errors are painful. So this is
13:59 one where the agent picks up the wrong
14:01 tool and whether or not it gets the job
14:02 done right, it's a tool it should never
14:04 have picked up in the first place. This
14:06 is especially common when you
14:08 incorrectly frame tools in the system
14:10 prompt or you don't make them available
14:11 in the harness in the correct way or you
14:13 have too many of them or they're too
14:15 long. Tools are something that probably
14:17 deserve an entire deep dive on their
14:19 own, but I will say here that the
14:22 ability to diagnose tool problems is one
14:24 of the markers of an AI fluent person.
14:25 And part of how I know that is that the
14:27 Claude certified architect program,
14:29 which recently launched, tests for this
14:31 failure mode specifically because it's
14:33 so important to building sustainable
14:35 agentic systems. And if you think, oh,
14:37 what's Claude certified architect? It's
14:39 nothing. Accenture is rolling this out
14:41 to hundreds of thousands of people. is
14:43 going to be like an AWS certification
14:44 shortly. Everyone's going to need it.
14:46 Here's another failure. Cascading
14:48 failure rate. So, one agent's failure
14:49 propagates through the chain. You never
14:51 had correction mechanisms and now you
14:53 have a failure of the whole run. That's
14:54 it is correctable if you put in loops
14:56 and verification in the right places.
14:58 The most dangerous failure of all I kept
15:00 for last. It's called silent failure.
15:02 It's where the agent produces a
15:05 plausible output and it looks right, but
15:07 something went wrong and the actual
15:09 result isn't acceptable in production.
15:11 Those are very difficult to diagnose
15:13 ahead of time and once you find them,
15:15 they're hard to root cause because they
15:18 tend to look identical to correct output
15:19 by most measures. I'll give you an
15:21 example. Let's say you're trying to
15:23 recommend a particular product to a
15:25 customer and it's brown leather boots
15:28 and the AI system comes back and says
15:30 it's recommending brown leather boots
15:31 and the customer is unhappy and leaves a
15:33 nasty review and something went wrong.
15:35 You go back and you see, okay, it said
15:36 brown leather boots in the chat. That
15:38 looks correct. The metadata on the
15:40 product says brown leather boots and you
15:42 have to dig and dig and dig and dig to
15:45 see that the issue is on the warehousing
15:46 shelves and someone actually shipped
15:48 blue leather boots and there are blue
15:51 leather boots pictured in the last
15:53 picture of the rotating carousel on that
15:55 skew and there was a mixup and in this
15:57 case it may be the agentic interaction
15:59 with an incorrect initial data set that
16:01 caused the problem but it still shows up
16:02 as a silent failure. That is the kind of
16:04 hard work that you have to do to get
16:06 these systems to work well. Now, if
16:07 you're an S sur, if you're a risk
16:09 manager, if you're an operations leader,
16:11 you already think in these failure
16:13 modes. This is not a big jump. If you're
16:15 someone else and you're just not used to
16:16 thinking in failure modes, I promise you
16:18 once you get into it, it's a little bit
16:19 addictive because it's like looking
16:21 through a puzzle and saying, "Where's
16:22 the missing piece? There's got to be a
16:23 missing piece in here." So, it's
16:25 absolutely something you can learn. Now,
16:27 once you understand these systems pretty
16:29 well, the higher level skill, again,
16:31 something I am rooting directly in job
16:33 postings is around trust and security
16:35 design. Basically, how do you know where
16:37 and when to implement these systems and
16:39 where and when to put humans in? Where
16:41 do you draw the line between human and
16:44 agent? Where do you authorize an agent
16:46 to take an appropriate action? And how
16:47 do you know the authorized agent only
16:49 took those appropriate action? How do
16:52 you keep an agent on guardrail so you
16:53 know it does not say something
16:56 inappropriate to a customer? So this is
16:57 a case where you basically have to build
16:59 the containers or the guardrails around
17:01 the agentic system in such a way that
17:02 you are confident that it will
17:05 predictably and reliably yield value in
17:08 production systems. This is a very
17:09 difficult skill because these systems
17:12 are probabilistic and just telling it in
17:14 the system prompt hey be good be nice is
17:16 not going to be good enough. So digging
17:18 in if we look at subsklls here you have
17:20 to understand cost of error right you
17:22 have to understand what is the blast
17:24 radius of particular problems the art of
17:26 building these systems and guardrails is
17:28 the art of saying what is the worst
17:30 thing that could happen let's get clear
17:31 on that and then work backwards because
17:33 you're never going to be perfect like a
17:36 misspelled email draft that's not great
17:37 a incorrect drug interaction
17:39 recommendation is potentially
17:41 catastrophic for the company and so you
17:43 have to understand where to put that
17:45 valuation and how to make sure that you
17:46 get the big things, right? Another one
17:49 is reversibility. Can you make this
17:51 mistake go away by reversing it? Now,
17:53 you can review a draft before sending
17:55 it. You can't necessarily review a
17:56 transaction that's a wire transfer
17:58 that's already gone out. That's gone.
18:00 Frequency is another way in which you
18:01 understand the risks of the system. If
18:03 it happens 10,000 times a day, it is
18:05 potentially a much bigger risk profile
18:07 than if it happens twice a day. Then
18:08 again, if it's twice a day going to
18:10 100,000 people, maybe you have to think
18:12 about it. This requires a depth of
18:13 understanding of the system that allows
18:15 you to really map customer impact
18:17 clearly and precisely. Last but not
18:20 least is verifiability. Can you verify
18:23 that this is correct? It's a big word in
18:25 this discourse and you have to look at
18:26 all of the answers you're getting and
18:28 you can't just be tolerating semantic
18:29 correctness. Semantic correctness is
18:31 when the LLM says something to a
18:33 customer and it sounds right. Functional
18:35 correctness is when the LLM says
18:38 something and it is right. Like an LLM
18:40 can say, "Hey, this is the right credit
18:42 card for you." And that sounds correct,
18:43 but if the credit card recommended is
18:45 the wrong credit card, it's still a
18:47 disaster. You have to be functionally
18:50 correct, and you have to insist that you
18:52 measure systems against that standard.
18:54 And so, a lot of what these job postings
18:56 tend to look for is people who have that
18:59 insanely high bar on quality and insist
19:00 on building systems that uphold that.
19:02 Now, let's say you've gone through this
19:03 whole process. You can build aic
19:05 systems. You understand the boundaries
19:06 they draw. Do you understand how to
19:08 specify intent? All the skills I've just
19:11 described. The crowning skill is context
19:14 architecture. How do you build context
19:18 systems that enable you to supply agents
19:20 with the information they need on demand
19:23 to successfully run at scale? This is
19:26 the 2026 version of getting the right
19:28 documents into the prompt, which is what
19:30 we were doing in 2024. So, you have to
19:32 understand what is persistent context in
19:33 your system. What is always there? What
19:35 is per session or per run context that
19:37 the agent needs? How do you make that
19:39 available? How do you make sure that the
19:41 data objects in your space are easy to
19:44 find and easy to traverse by AI agents?
19:46 How do you make sure that there isn't
19:48 dirty and polluting data that confuses
19:50 the AI agent in your context available
19:52 to be searched? How do you differentiate
19:54 between what is pulled into context and
19:55 what isn't? And how do you start to
19:57 troubleshoot when agents start finding
19:59 the wrong context? Context architecture
20:02 is one of the hardest things to do in
20:05 2026 and it's something that many
20:07 companies are now willing to pay almost
20:09 anything for. If they can get this
20:10 right, it enables them to not just build
20:13 one agentic system but to build dozens.
20:15 It's a massive unlock. And the people
20:17 who can think through the data side of
20:20 things logically and put that in front
20:22 of an agent in such a way that they can
20:24 verifiably show that the agent can do
20:26 the work, those people can write their
20:28 ticket. And you know what? You don't
20:29 have to be an engineer to do this. If
20:32 you're a librarian, if you are a
20:34 technical writer, you have a lot of the
20:36 bones of this skill. You have the
20:37 ability to understand technical
20:39 information and where it's filed and
20:41 where it goes. In a sense, context
20:43 architecture is like building the Dewey
20:46 decimal system for agents. You have to
20:48 understand how to build a library that
20:50 an agent can easily search through and
20:51 find and say, "Ah, this is the right
20:53 book. I have to pull this for this job."
20:55 And you're doing that with company data.
20:57 And that is a skill that you can test
20:58 for, that you can hire for, and it is
21:01 highly in demand right now. Okay, last
21:03 of the seven skills. This seventh skill
21:05 is on almost every senior job posting.
21:07 It's called cost and token economics.
21:09 I'll simplify it for you. Is it worth it
21:11 to build an agent for this job? You have
21:13 to be able to go through and calculate
21:16 the cost per token for a given task and
21:19 reliably say, if I put an agent against
21:21 this and it burns 100 million tokens, I
21:23 can prove this is worth doing or I can
21:25 prove it's not worth doing. And I can do
21:27 that ahead of time before I put a bunch
21:29 of agents against this. If you don't
21:30 know how to do this, and in particular,
21:32 if you don't know how to do this in a
21:34 world where you have model choice, where
21:35 you can pick your tokens, where you have
21:37 to pick the right model to pick the
21:39 right tokens for the economics of the
21:41 task, and you recognize that all of
21:42 these models are changing their pricing
21:45 all the time. That's the challenge. That
21:46 is what you need to be able to do. Like
21:49 imagine a world where token cost as a
21:51 whole is falling very rapidly, but you
21:53 may need frontier model pricing for
21:55 certain tasks. How do you ensure that if
21:57 you're being tasked with getting a job
21:58 done, you can get the right mix of
22:00 models on the job, you can calculate out
22:03 the blended cost of the task, you can be
22:04 confident that you're paying the right
22:06 amount, you're getting ROI on the task.
22:09 That is a senior level qualification.
22:10 You're not surprised, right? It's highly
22:13 in demand. Being able to do that is
22:15 basically just applied math. It's it's
22:16 actually you can actually build
22:18 spreadsheets and calculators that help
22:20 you to do this where you can just change
22:22 variables and say I think it'll be a 100
22:23 million token task and you can see
22:25 immediately across six different models
22:26 how much it would actually cost assuming
22:29 a given weight. And it's actually not as
22:31 hard as you would think to figure out
22:32 what those different components would
22:34 cost because you can put together a
22:36 little prototype and you can very easily
22:38 cycle through tokens, see that it's
22:39 plausible with this model and not with
22:41 this model. see roughly how many tokens
22:43 it takes across three or four runs and
22:44 you can start to build a plausible
22:46 model. This is a situation where it's
22:47 high school math, but you're getting
22:49 paid like a senior architect or a senior
22:52 engineer because you're fundamentally
22:53 taking those mathematical skills and
22:55 applying them in a very fluid and
22:56 fastmoving world to help the
22:58 organization be very costefficient with
23:00 these agents which are not cheap to run.
23:02 Like if you're burning through a billion
23:03 tokens with an agent, you'd better be
23:05 sure it got it right. It better be worth
23:07 it. If you're wondering what kind of job
23:08 titles did I look through for this? Is
23:10 it only engineering? The answer is no.
23:12 There are operations titles that have
23:13 these skills. There are engineering
23:15 titles that have these skills. Yes,
23:16 there are architecture titles that have
23:18 these skills. There are product manager
23:19 titles that have these skills. There are
23:21 AI reliability roles that have these
23:23 skills. People are calling them
23:25 different things. And what I have done
23:27 is dig underneath and map out seven
23:28 common skills. And we are going to see
23:31 in 2026 more and more new skills
23:32 emerging and more and more new jobs
23:34 emerging because fundamentally we're
23:36 rebuilding job families around agents.
23:38 And so you're going to find that someone
23:40 is going to be very clear about wanting
23:42 someone with high specification quality
23:44 to be clear about intent with agents at
23:46 initial parts of the run or someone
23:47 that's going to be really good at evaluing
23:52 skill sets are the underlying skill sets
23:54 and part of why I'm confident they're
23:55 not going anywhere is that they're tied
23:57 tightly to how AI actually works. It's
23:59 like the agent may get 10 times better
24:01 at doing complex longwriting tasks but
24:02 you still got to have an email at the
24:04 end. You still got to specify your
24:05 intent. Make sure it's like this is what
24:07 we're going to do. You still have to be
24:08 able to search your context
24:10 appropriately. These these skills are
24:12 skills that you can bet on. These skills
24:14 are skills that companies are betting
24:16 careers on and they're desperate for
24:17 them and no one can find them. And look,
24:19 if you got to the end of this video and
24:20 you're like, "That's me. That's me. I'm
24:22 raising my hand." Head over to the job
24:24 board that I'm putting up here. Go check
24:27 it out. Go put up your profile and let's
24:29 get you into the mix as part of a vetted
24:30 pool of talent so that we can simplify
24:32 this job finding and hiring process in
24:34 the HI. If that's you and you're like,
24:36 I'm a hiring manager and I've got to get
24:37 these kinds of people, same thing. Head
24:40 over there. If you're like, I want to
24:42 get there, that's why I'm creating the
24:43 guide to help you get there. That's why
24:45 I'm working through a course with you on
24:46 the Substack where you can actually go
24:48 through and teach yourself these skills
24:49 and you can self diagnose and say,
24:51 "Okay, which one do I need to work on
24:53 and how do I get better at it?" So, my
24:54 goal here is to be practical. I want
24:56 this to be something that is
24:59 distinguished from other AI self-help
25:01 guides by being specific enough to be
25:04 useful, by being grounded in actual job
25:07 posts, by being grounded in the skills
25:09 that I see employers looking me in the
25:12 eye and begging me for after hundreds of
25:14 interviews they can't find them. I have
25:16 seen people throw up their hands and
25:18 tell me you go out and interview people.
25:19 I've interviewed hundreds of people. I
25:21 cannot get this job roll filled. That is
25:23 what it's like on the other side. And so
25:24 if you want to be the person who is in
25:26 demand in that world, this is for you.
25:28 Bookmark this video, go back through it.
25:30 I know it was dense. You can feed this
25:32 transcript to an AI and ask it to
25:33 explain it to you. People do that all
25:34 the time. They tell me so in the
25:36 comments. You can get help to get this
25:38 done. I hope this has been helpful. Cheers.
