$75M founder reveals his Agentic Engineering setup

0:00 Everything you ever learned about building software is kind of thrown on its head. >> AI makes high agency intelligent people look really good. It makes dumb [music]

0:07 low agency people look worse. >> The percentage of time that you spend on engineering the markdown should be higher than the percentage of time that you spend on actually executing the

0:17 code. >> At a point very soon, no one including all engineers will read code or write code. >> This is Alex Lieberman, the co-founder

0:25 of Morning Brew and 10X. and he's joined by Dan, the director of engineering at 10X. In this podcast, we talk about how any company can become AI native, why markdown files are worth more than your

0:37 code, and what is the last remaining mode that AI won't kill. If you want to see how people on the cutting edge of AI actually work, watch until the end. This is the David Andre podcast. Enjoy. Alex

0:48 and Dan, I have a question. How do you take a company that's old and slowmoving and make it AI native? I think the answer to the question is kind of like how do you take anyone who has a set of

1:00 existing context and get them to do something different and the answer is meet a company where they're at, where their systems are at, where their people are at, where they are in their AI

1:09 transformation journey. If you were to reimagine what a consulting business for Fortune 2000 companies looks like in 2026, it has to be a full stack partner. Like historically, some of these biggest

1:21 consultancies in the world only did one thing. They only would basically be told a problem. You would have a group of consultants and their manager work on a solution to that problem as a slide deck

1:33 for a few months. They'd meet up with the client. They would deliver the solution via this 150page deck. They'd give them the deck and they'd say, "Okay, peace out. See you guys later."

1:42 And then the client would would rely on having to solve the the problem themselves. Our view is in a postAI world, you need to touch every part of the value chain in a company's

1:52 transformation. Developing their strategy, doing non-engineering work to drive change management and everything that needs to happen around the tech to actually make the tech valuable and then

2:01 actually forward deploying into a company. to to your question, there are a lot of companies, a lot of Fortune 500 businesses where the first thing we do is literally taking their seuite or

2:12 their executive leadership team and teaching them how to use clawed co-work or claude code. And then for other companies who come to us and say, "Hey, we want to build back office agents for

2:22 invoicing, AR, AP, HR, etc." We go and we meet them where they're at with that problem. So really, it's like how do you decrease the friction to as close to zero as humanly possible?

2:31 >> I see. So I guess that's one of the like biggest separators of like an average consultant where he like gives you something and then you have to figure it out and a great consultant that like you

2:40 know helps you implement it. What are some other things because a lot of people want to be an AI consultant you know if people tried starting AI consultancy in 242.

2:48 >> Yeah. >> What are some other separators? >> I would say Arman my co-founder and co-managing partner myself we did not dream of being consultants like neither

2:58 of us are consultants. Our actual vision for 10X is we want to build a modern-day Bell Labs and for different reasons. I have realized about myself if I'm not home with my wife, my daughter doing

3:12 something else, it has to be worth it. And to me, worth it means solving really hard problems that are on the frontier of technology with the smartest, most driven people I've ever worked with. For

3:22 Arman, I think it was in high school or college, he read a book called The Innovators. the innovators. It's an exceptional book. Talks about the history of computing literally from

3:30 Charles Babage and Ava Llace all the way to effectively AI. And in this book, it talks about the arc of innovation and what were the raw materials that made kind of these centers for innovation

3:43 successful. Whether it was Bell Labs with the transistor, whether it was DARPA with the internet, whether it was Xerox Park and the short answer is the entire time it is effectively brilliant

3:53 people who are given a great deal of autonomy but they are given direction, distribution and resources. And so what we basically said is if we want to have all five of those ingredients and we are

4:05 we do not desire to raise hundreds of millions of dollars from the start, how do we do that? And elegantly, the answer starts looking like helping the biggest companies in the world solve their

4:15 hardest problems with AI. By solving these problems, enterprises are telling us their biggest challenges that they're that they're navigating within their business. They are paying us to do so.

4:24 We so we are cash flowing the business. So our view is in helping enterprises end up on the right side of history. We are getting all the raw materials we need to effectively build an applied AI

4:34 lab where we incubate some of the most consequential technologies in a postAI world. Let's see. Seance 2.5 just came out, but most people are going to be paying for every generation. On top

4:44 view, however, Ultra users get Cance unlimited for 60 days. And unlimited is way better when you have a workflow that can actually use it. That's what top view canvas is. It's a visual workspace

4:56 where an AI agent plans and builds your entire video. You drop in a product, a script, or just a brief, and the video agent plans the storyboard and generates the scenes. Everything lands on a visual

5:08 canvas and every step stays editable. And with Cedance 2.5 under the hood, you get native 30 second singleshot videos. Top view canvas can also take up to 50 reference assets. So your product or

5:21 characters stay consistent across shots. And if one section comes out wrong, you don't need to worry because the clip edit feature can regenerate just that part. And once you have a workflow that

5:32 works, save it as a skill so you can rerun it for the next product or spin out multiple ad variations from the same setup. And remember that 60-day unlimited Cense 2.5 deal is for ultra

5:44 annual users. So click the first link in the description and give Top View a shot. One of the most interesting roles you could possibly have as an engineer is this kind of like applied AI forward

5:55 deployed engineer because you you can accelerate the speed to solving very hard custom fit problems for enterprises and enterprises previously could only have these oneizefits-all solutions. Now

6:08 you can go in and and kind of find the the gnarliest problems possible in enterprises and solve many of them um in kind of this very rapid succession. And so if you can reduce that time to magic

6:20 very early on in in one of these engagements and just show the seauite or whoever the stakeholder is um how much leverage their teams and their orgs can get with AI, it it becomes this cascade

6:32 of everyone else in the organization kind of wants in and and we get to um build more and more for them. >> So it's basically like innovation as a service.

6:41 >> Yeah, that's exactly right. >> I think consultancy is like underelling it, you know. I think that's totally true and I think that works to our advantage because honestly I think

6:49 unfortunately the best technical talent has never made its way to consulting because like one way to sell consulting is you go into a company, they tell you a problem, usually the answer is firing

7:00 people, you deliver them a deck, you walk away, you never see the fruit of your work and then you move on to the next client. Or you could position consulting as a huge company that has

7:09 messy, complex, and really important challenges to solve. They don't have the answer and they need an a team of people to basically provide them innovation as a service and not just innovation as a

7:21 service but innovation and implementation like actually driving ROI and growth through the insights they provide. The cool thing about it is you don't have to just do this with one

7:29 company forever which is often times a thing that I think uh creates boredom or plateauing for engineers in any org is like you work on one very specific challenge for a very long time.

7:40 Consulting by design offers you diversity of work. >> Yeah. So you know that ADHD is satisfied and you still work for the same company. You don't have to career hop, but you

7:50 have different challenges. >> Yeah. Exactly. And there's a few common questions that when we talk to a client, they ask us and we need to provide thoughts on. One question is is like how

8:00 should we even be thinking about where there's opportunity for AI in our business? Like again, we meet people at different parts of the journey. And I would say half the time it's a seale

8:10 person reaching out to us saying, "Guys, I'm in this group chat with a bunch of sea friends and they're all talking about all this cool AI they're doing and I have FOMO and I don't know

8:20 what we should do with AI, but we need AI. Can you help me with AI?" And one of the initial frameworks we provide is this idea of single player versus multiplayer AI. And so the idea is that

8:29 single player AI is if I was like, "Hey David, uh you traditionally would use Google to put in a query and get an answer. I'm going to put GPT or uh Claude in front of you. Ask that same

8:43 query and see what you think about the answer." Zero behavioral change. >> You just do that. And that is valuable. And and in the context of an enterprise, what that looks like is an enterprise

8:52 and this is usually the first step that enterprises take in their AI transformation journey is they sign a large token agreement uh with one of the labs. So e either with open AI or with

9:03 anthropic they're like hey we'll commit to510 million worth of token spend and we just need to make sure that we have you know secure instances of using uh codeex and

9:16 GPT or claude code and co-work at our company and the first step of single player is literally just putting the power of a state-of-the-art model in front of your people and then ideally

9:27 teaching them how to actually use these things >> in their day-to-day workflow. That is really valuable and I would actually argue for most companies that is where

9:35 most companies should start but that is single player like people are only getting leverage from use of the technology for their own work but there is no compounding effect. Multiplayer AI

9:46 is this idea of where you look at horizontal technologies or processes in your business where by reinventing them or reimagining them you're not just creating leverage for one person you're

9:56 either creating leverage for the entire company or an entire function. So that could be everything from building uh you know a an SDR or sales agent for a sales or that helps every additional every

10:09 seller get back time that they were uh spending before on like the logistics and the minutia of their job versus actually talking to clients or we say this all the time most clients come to

10:20 us with an AI problem and it turns into a data problem where we're built we're doing some form of data engineering for clients because they are nowhere close to being data ready to actually build

10:30 any form of agents or solutions on top of again another example of multiplayer AI where if you get that right the leverage created for the entire organization is massive and so our

10:40 framing is we do both single player and multiplayer AI at 10x but where you're really going to get exponential benefit and where it's going to be really hard and painful and messy for you and where

10:51 we can be most supportive is on the multiplayer side. >> Yeah. I want to share a tip we do uh at our much smaller company is that like I have everybody work at different GitHub

11:00 repos right so we have one for like hiring one for brand deals one for YouTube videos and then anything you do instead of just chatting in the cloud web app or chat GB web app I force

11:10 everybody to chat in codex and cloud and you know save their work in these markdown files in these repos so that anybody else they can load up their own clo with fable and you know even if

11:21 that's not their area of business that they work in they already have all those project specific skills, all those markdown files, all that context and you know it's it's a complete game changer

11:29 because even people who you know half a year ago were not technical and not really using AI as much as they could now they do and you know sharing those skills is really it's really magical

11:39 because somebody who doesn't work in that they they feel like they can do it because you know they have their own fable fable reads those files gives them the brief and now it's like a form of

11:48 multiplayer. >> Totally. And one one thing I'd just add and we can uh up to you whether we talk about this now or later on, but I think like this is exactly how Dan and the

11:56 engineering team at 10X think about like our entire SDLC and this this idea of context as code. Like it's interesting because engineers have long thought of being really organized and methodical in

12:11 the way that they structure files and information when building software. In traditional knowledge work, that is not the case. So I think actually in a lot of ways non-technical knowledge workers

12:22 are playing catchup to get to a place how engineers have always thought about kind of like setting up the scaffolding of how they work. And I think with engineering at 10X we're just taking it

12:33 to another level of h of how kind of careful and prescriptive we are with the context we provide to coding agents that we work with. >> Yeah, we can touch on it a bit more. Go

12:42 ahead, Dan. Yeah, I mean I can I can go into the demo a bit later, but I would I would think of it as this new kind of problem where previously you were solving this coordination problem

12:51 between just humans and humans. And so the way that like agile development worked was just a ton of ceremonies. You would meet every morning, you would talk about it, talk about the work, and then

13:01 you would have retros at the end of every week and you just fill up the calendar with as many meetings to to convey information as possible. But now it's really a coordination problem

13:11 between humans and other humans and agents as the third party. Um, but agents are kind of like in this main character spot where they're the ones actually doing the the work on the

13:21 ground. And so I think that it is like so much more important now than ever to be a written first culture and kind of what you're saying David uh along the lines of uh storing up all this markdown

13:34 so that uh both humans can get on the same page about what they're working on across different projects but also so that agents can be very quickly briefed on um what's going on. And then like

13:46 Alex said at 10X, we're kind of taking this um really far to the extreme where if you think through from first principles like how would you redesign the SDLC for agents? Agents can maintain

13:59 coherence across so much more information at the same time than humans ever could. And so you can structure like bigger and more complex blueprints that lay out uh more of the

14:10 architecture, more of the plans um so much further in advance than you ever could before and and then you ever could in these like agile humanto human ceremonies. And so we have these very

14:24 concrete structures and types of uh markdown artifacts that we write for all of our projects and and structure it for everything that we're working on so that agents can easily validate. I mean, kind

14:38 of like Alex is saying, we treat this context as code, and I'll show this in a bit, but agents can immediately pull all the right context at the right moment and kind of keep all this in sync as a

14:50 new form of um like software machine, so to speak. >> Yeah, I think uh with each project I start like I find myself keeping more and more stuff in the codebase like

15:00 literally I have like /doc/marketing. I have a you know user feedback file where any user feedback I just put there and it's markground files. I have a /doc external all of my you know deployments

15:11 and stuff that's external to the codebase is stored in there because again who knows if the agent needs it you know maybe there's like some uh deployment issue with the front end

15:20 hosting if he knows what what it's configured like what what variables are there maybe he finds a missing environ we need a new word for it because it's not a code base anymore you know it's

15:32 like in the past it used to be just code but now it's it's kind of like a context base >> we can we we should we should call it something like a meta harness.

15:39 >> Meta harness. Okay. >> Yeah. Yeah. That that's what we called it internally at 10x. And I think the like it's what are we trying to achieve? It's like coding agents offer speed that

15:48 was never possible, but the longer running the task, the more that you risk entropy and diversion from plan. And so the question is is how do you minimize entropy and diversion while optimizing

15:59 for the speed that is kind of the power of this technology? Okay. I also want to talk about the content machine you guys have because this is one of the most unique systems and I think we will see

16:07 more and more companies doing it where not just the founder and you know the marketing guy but like more people are creating content. So tell us more about that.

16:15 >> I mean and this will definitely resonate with you given that you kind of have a foot in both worlds both like in building software but also obviously you you've seen the power of content

16:25 yourself. Like one of our let me start with like one of our highest level thesis at 10x is that in a postAI world the number of moes remaining in business is shrinking like I think

16:38 what what are moes today it's like network effects to some extent data to some extent I would argue people people think this this is kind of I think people is more of a moat than ever

16:48 before because AI is just like AI is a funhouse mirror it makes high agency intelligent people look really good it makes dumb low agency people look worse. And so I think people is a moat than

16:58 ever before. And then I would argue trusted distribution is more remote than ever before, right? Like you you build a piece of software and likely there are literally millions of other engineers

17:11 that can build that piece of software pretty quickly. Now what is going to earn you dollars from from an audience? It's the fact that you have people who trust you and that trust gets turned

17:22 into dollars in the bank. And so we've always had this view that we're building a media company on top of 10X. And this media company is going to be a combination of like brand driven content

17:33 like newsletters and deep uh editorial that is both technical and non-technical but then also talent driven content where we're literally hiring David before David's became big. That's like

17:43 we're bringing on like dozens of those to our business. But we also want to turn our employees at 10x into creators. And so the thesis the the the question I asked myself was like how can I turn

17:58 engineers or strategists at 10x into content creators and starting with textbased content because I think it's just way easier to leverage AI for that versus video and multimedia right now.

18:10 And same thing with myself, like I don't have all day to create content, but I started as like the first best marketing channel for 10x. And so I need to make sure that I kept my content output

18:21 going. And so the idea of the content machine was how do I build a system that allows for content abundance that enables full-time people to still create if they have 30 minutes per day, but is

18:34 not an AI slop cannon. That was the prompt. What I can do is why don't I share my screen and kind of take you through what the system is and then feel free to just ask questions throughout.

18:46 So basically I I wrote this post on and I think it's just a good overview and then we can I can actually run the machine if we want while we're talking through things and I can even have it

18:54 run when Dan goes through his stuff but basically the way the content machine is set set up is there's a process layer and then there's a personal layer. The process layer is everything that lives

19:04 in git for my team. And that is basically the p the pipeline like all of the steps that the content machine goes through to take uh almost like the assembly the the the factory line from

19:15 need I need an idea of what to create content about to finished idea that it has been distributed on X LinkedIn or another platform. And then there's the personal layer which is the files that

19:26 live on your computer which is basically the cu uh codifying of your voice content lessons that you've provided to the content machine in the past and any other uh important information about you

19:39 that informs the way the content machine um outputs in a way that is representative of you. But the one big thing I'll share about how you can create

19:50 an AI system that puts out content that isn't slop at a time when I think the models are still like pretty bad at writing like like they're they're good but they always

20:04 stink of AI. >> Yeah. The same patterns. >> Exactly. And and I think like you know I've seen people create these skills like this guy Peter Yang created like a

20:12 no AI slop skill that like takes all the tells but I think with every change in models there's going to be new tells and it's going to be kind of this constant game of cat and mouse chasing the new

20:22 tells. But my view is is the way you can guarantee no matter how good the models are um that you can guarantee it not being slop is that the AI kind of helps with everything other than what the

20:36 human is needed for obvious and what I would argue the human is most needed for. Not just in a content process honestly in like most processes even with uh building software is the first

20:47 mile and the final mile. >> Yes, I call it the origin. I actually I actually made a short about this. I said like the the number one difference between like AI content that's trash and

20:56 AI content that is valuable is if the origin is human. >> Exactly. Exactly. >> Real quick, you can now try Deep API completely for free. In fact, we're

21:04 giving $1 of free credits to anybody. So you can try it yourself. So let me show you what it looks like because deep API I would say is the single most powerful API key you can give to your agent. So

21:14 let me show you. I can do like something use deep API to scrape 50 CTO's of AI startups valued at over $1 billion or more and give me at least two forms of contact for each output as table and

21:33 deep API will scrape their social medias different you know their Twitters whatever it takes to find these people all with a single API key whatever you're doing with AI agents deep API

21:44 makes them a lot more powerful and gives them access to worldclass deep research, powerful, fast, affordable web search and scraping for any website or any social media. And now after building it

21:54 for two months in stealth mode, we're opening up deep API to the public and you can try it completely for free. Just create an account, you get $1 of free credits, set it up. The installer is

22:04 very simple to run. It installs a skill on your computer. So you can use it with Codex, Cloud Code, Hermes, or any other agent. And yeah, feel free to use it. um test it for your use case and watch your

22:13 agents become a lot more powerful just by giving them a single API key. Try Deep API by going to deepappi.co or click the second link below the video. And so what I would do if there

22:24 are steps in a content, right, if you think about what a content process is, and this is how I built this whole machine is I literally drew out on a piece of paper what is every single

22:32 single step in the content process. And it's like step one have to think of an idea. Step two, I thought of an idea. Step three, I research the idea. Step four, I get all of my thoughts out. Step

22:43 five, I turn it into a cohesive draft. Step six, I edit that cohesive draft. Step seven, I'm pissed. I don't think it's as good of a draft as it should be. I rewrite it. Step seven, I finish the

22:55 draft. Step eight, I post it. And so, like, I wrote this out and I was like, let's rebuild this to be AI first and keep the human where the human needs to be. And my view is the human always

23:05 needs to select the idea that they want to create content about. They need to provide their own words about this content idea. So if I was to say, you know, hey David, what are your thoughts

23:18 on Kimmy K3 to actually have you create a good piece of content around it? I would want to interview you for 30 minutes after you've done research into the model, what makes it kind of like

23:28 the elegant solutions to problems uh around memory that it's solved. You would give me your thought for 30 minutes. Those thoughts become effectively the the transcript or the

23:37 markdown file that the entire content machine operates on. And basically the goal of the content machine is to give you ideas to to pick from and then take the words that you've given and not

23:49 change your words but just make sure your words flow, that they transition well, and that you have an editor who's checking what you've said. to add in places where you need to provide more

24:01 context because you didn't share it with your initial thoughts or where you need to make other tweaks. And so this is basically the process is in the content machine you do creator select first

24:11 because all of our uh employees at 10X have used the machine. You first select like who you are which then loads in the files that are specific to you around your voice and your content lessons.

24:21 Then we have this thing called the oracle which runs and the oracle scans your last seven days on slack, notion, gmail, linear, git and it hunts for spikes and spikes are basically how much

24:33 point of view is there, how much story potential is there, how much emotional intensity is there, how much lesson or framework is there and how much depth is there. In addition to pulling ideas from

24:44 internal sources in your company, it also pulls from external sources. And I actually have I use slashl last 30 days which is Matt Van Horn's um skill uh for doing this. So basically there's a

24:57 social sweep across Reddit X, YouTube, HackerNews, etc. on the topics that you think you want to create an idea around. This basically produces a list of like 15 ideas. All of those ideas get added

25:08 to a notion database that is created whenever someone spins up the content machine for the first time. It's called the vault. And so, you know, let's just say I go through the content machine and

25:18 I pick one idea. The other 14 ideas may not be bad. I just don't want to use them now. They get added to this database. Then, let's just say I pick the idea around Kimmy K3. Let's say I

25:28 like, you know, I'm not deeply technical, but I want to have a perspective on it. I run I could do an intermediate research step where it researches the the um original white

25:38 paper. It also tells me what are current developments, what's already being said in market, what are contrarian angles, and what are open questions. only I as Alex Lieberman can answer. Once I have

25:47 my thoughts, I have an interview panel interview me. And so this is probably the most important step in the whole content machine because this is the idea of instead of Alex interviewing David

25:57 and like a human needs to interview a human, I have basically these six personas, Tim Ferrris, Joe Rogan, Larry King, Howard Stern, Michael Barbaro, and Barbara Walters each as a different

26:09 skill. So like their way of questioning is codified in six separate skills. They ask me questions one at a time. They force me to be specific. Anytime I have a vague answer, they push back. We

26:21 probably go for 20 minutes. And the way they ask me is over text in clawed code. I yap to text my answer. That gets turned into the full markdown file with the transcript, the key stories, the

26:32 core insights I share, the quotable moments. Then I tell it what format I want it written in whether it's LinkedIn post, Xthread, long post, playbook, podcast promo, etc. It finds the skill

26:46 which is my codified version of how have I written let's say it's a LinkedIn post. How have I written LinkedIn posts in the past and what have been my best performing LinkedIn posts and what was

26:55 the format of those? Right? It takes the raw markdown file. It refineses it for flow but not changing my words into that format. >> Then I have a writer's council or an

27:06 editor's council. It's the same idea as the interview panel but for editing. So Morgan Hel, Tim Urban, Sean Pur, Greg Eisenberg, David Prell, and a slot detector. They go through and they score

27:17 it. If it is under a nine out of 10, it goes through a revision loop. If it's above a nine out of 10, it's done. And then I get the piece and then I can run it through a repurposing engine where,

27:29 let's just say I finish this long form piece on my like my takes on Kimmy K3, why it's important, what people should know. I can then have the repurposing engine turn it into 10 pieces of

27:39 derivative content, but it goes through the same process where it pulls the skill for what is my my way of creating an X article, what is my way of creating a LinkedIn article, it creates it and it

27:49 runs it through the writing council again. So all these pieces go through the same flow and then there's a distribution piece where actually I get a UTM tag for each. I can autopublish it

27:59 from the machine and we can uh monitor performance and the whole idea is by monitoring performance then the machine gets better in the future of writing post because it knows what top

28:09 performers are. And then the final piece is at the end of every session I provide feedback on what I think the content machine did well with creating a piece of content and it turns that into a

28:20 lessons file. So every time it re write it it writes a piece of content in the future, it checks contentlessons.mmd to make sure it's not making the same mistake again.

28:31 >> And how many examples do you usually need in these like human written, you know, articles, LinkedIn posts until you feel it's good? >> Yeah, it's a good question. I So I don't

28:41 know the exact answer because I'm I'm a unique case where I have a lot of prior content like you know to train this. I had 350 podcast episodes. I have like 10,000 posts across social. You

28:52 definitely don't need that many. I would say, let's say Dan wants to set up the content machine. And let's say Dan has never posted on X or LinkedIn before. He basically has three ways for how he can

29:05 train his voice for the content machine. Way one is there's actually like the content machine will run a voice interview where it will interview him to get his voice out and turn that

29:17 interview into his voice file. That's number one. Number two is he can take basically feed anything that shows his voice, whether it's past text messages, emails he sent, or Slack messages, that

29:27 turns into his voice file. Or the third is, and I've recommended this to people, is just start with me as the persona. Literally have it use my voice files and then over time as you give it feedback,

29:38 it'll just naturally morph into how your you want your voice to be different from my voice, but at least you have a voice you feel comfortable with. Dan, what's your been your experience with this

29:47 content machine system? >> I mean, it's great. I think that it is nice to start with Alex's voice because he's trained uh his content machine or his voice on so much history of what

29:59 works well and what doesn't on on social and then kind of around the company everyone's just whisper flowing their um takes about whatever the latest model is or or trend in um agent engineering and

30:13 uh just just translating it and getting it out. And I think since Alex kind of rolled out the content machine internally, we've seen this huge boost in the amount of people at 10x actually

30:23 posting content. Um, and I expect that to just increase uh over time. >> And David, two other thoughts here is like my view is it is generally really hard to get people whose full-time job

30:34 is not creating content to create content. And I think the way you need to do it is you have to decrease friction and you need to provide incentive. So content machine was decreasing friction.

30:44 And I think we can decrease friction even more. So what that could even look like is, you know, how I was saying >> Yeah, exactly. It's like we're we're hiring full-time creators in the future.

30:55 I could imagine our full-time creators, they partially act as producers for our team where they go they record conversations and just feed it into the content machine. Or imagine Dan's on an

31:05 inter like we're uh in a few hours uh we're talking about a piece of internal IP that we're building at 10x right now. if that whole meeting gets recorded and then we just run the content machine on

31:17 a hook where anytime there's a new meeting recording in notion transcripts, it just autocreates the piece of content. And then the second is incentive to basically in my mind go

31:26 from cold start of no one creating content to getting people to build a habit. We ran a competition where we were we gave away $5,000 over the course of a month to people to create content

31:39 and we ran different games where it wasn't just about whose content got the most impressions. It was also like whose content did the best job of educating an audience on a topic or whose content did

31:50 the best job of taking actual work we're doing with a client and storytelling it to the world. So I think you have to do both incentives and lowering friction to make this possible. Yeah, I think uh one

32:00 actionable thing we do is like when we have a Google meet, we have a noteaker there. Everything is automatically pulled through composio CLI and literally those transcripts like if

32:09 you're already discussing something whether it is you know the more senior people at the team or somebody coaching a new person that is like content opportunities right and then you can

32:18 have like a frontier model like fable analyze okay these things private super private never share but like these 10 things you could create tweets about them and that's another way we're

32:27 reducing friction where like you would have that call anyway but you don't even realize usually like the the internal conversations are the best content you know like you said if a model like Kim

32:35 K3 drops, what do you think about it? Are we really going to use it? Would our competitors are using it? Right? Like these kind of, you know, it's almost like this locker room talk, but like for

32:44 a company >> Exactly. >> where people are share their hot takes, but then again to open the phone, you know, write the tweet, it doesn't sound

32:51 as good. >> Totally. And and I think look, we're doing with this this with writing first. We also think there's a lot of opportunity to do this in video. like

32:59 one of our engineers at 10X CJ had like he's a longtime fan of Theo's content and he's been like I would love to do just like my live stream around like the five biggest topics in AI and

33:09 engineering right now with my perspective and I think the hard time right now the hard thing right now is AI is not really good at video editing and so you either have to have the internal

33:19 resources or you have to pick a format where you really don't have to edit but I think that'll be kind of like the next step of how do we build workflows that help on all mediums content, not just

33:29 text. >> You mentioned briefly Matt Van Horn and you had a nice interview with him and you said something which is a great hook by the way that one of the most

33:38 productive engineers you know cannot read code. So I've noticed the same pattern actually where you know some people who you would think they would be insanely productive with AI because they

33:47 like 15 years of experience in great software companies were actually very slow to adopt it. So tell us more about this pattern of people who cannot read code are usually more effective with

33:57 agents. >> Like first of all, do you think people should read code anymore? Like I feel like this was a very hot take by Matt. He's not an engineer, but he was he was

34:05 saying this as a generalized statement. And I was like I want to have a follow-up episode where I host a debate cuz I know there are going to be people who take the other side of the trade

34:13 here. But he was like I think in at a point very soon no one including all engineers will read code or write code. What do you agree with him or not? >> I definitely agree with the idea that

34:24 like the amount of code that you read is going down over time, but I think what you're outsourcing is the thinking and the processing over all that code, but you can't really outsource the

34:35 understanding. And so in order to structure the plans and the codebase and everything in the right way, I I still think and I actually like am more and more bullish on this over time that it

34:47 matters a ton um how deep you are in the fundamentals of engineering. And so I think that if you would take a an engineer who has really bad tendencies and and maybe pre-AII was not a

35:02 fantastic coder um but they really lean hard into AI that's actually going to like amplify those bad tendencies times 100. But maybe if you have someone who is uh less of an engineer, but is more

35:15 organized and um cares more about the structure and and knows how to pace things better and is just better at working with the agents, maybe they'll actually have better results than that

35:25 person who had the 100x negative outcome. But then I think the the best case scenario is someone who deeply understands the full stack of engineering and can um push the AI to to

35:36 its limit. So can you guys share your setup, your enchanting engineering setup at the next? >> Yeah, definitely. So um maybe as a a preamble, we already kind of talked

35:45 about the the meta harness and and the idea of constructing a a packet of context as code for every single engagement that we have um and for every project that we're working on. And so I

35:59 think the most important thing here is defining all the the sets of artifacts that the agent is going to be writing or that we're going to be exchanging between humans and agents. Defining the

36:09 the structure of those things and then being able to validate and and easily pull from those artifacts so that the agent can kind of self-correct. Um so I'm going to just go ahead and share. So

36:22 the the big pattern from what you're saying sounds like the thinking of like a architect is way more important than a specific like you know develop this feature or learn the syntax.

36:31 >> I do think that is true. Yeah. And and I think so one thing you'll see here um if we so we're in this workspace where we actually typically will separate our code repos from our um what we call our

36:44 project management repo which has all of these um has all of these artifacts in there. And so this file that we're looking at right now is what we'll call it's one of the types of artifacts. It's

36:56 called a convention art [clears throat] artifact which will have this con prefix. It has some metadata at the top so that the agent can um parse all of these artifacts and and find them. But

37:08 then it's an index of all the other conventions that we have um in the repo. And so this is the type of thing that will really keep an agent like on rails and following all the really good coding

37:20 patterns. And so I think it's super important that you know what you're doing in terms of engineering and and good patterns when you build up all these conventions files. But if you

37:29 build them in the right way and then you have this system to validate against the conventions, you you can worry a lot less about the quality of the code because you know that it's being

37:38 followed um all over the place. I'm just going to run um some commands to start to kind of show the CLI that we use to manage all these artifacts. So we're in this um workspace right now in in 10x

37:54 process and I'll just say um 10x context and right now I'll say the mode will be um for the operator which is for the human but we're we've designed this CLI to be both agent facing and human facing

38:08 and so right off the bat this is the type of information that the agent can just immediately get. It's going to report um about the artifacts that exist. So it'll have we have six epics.

38:20 This is one of our types of artifacts that defines kind of what we're building and the milestones to get there. And then 20 specs which are the the way more detailed technical documents about um

38:32 the exact architecture and ticket by ticket how we're going to build the thing. And then you can see we have 35 docs which includes those um convention documents. And then it's also reporting

38:43 what version of the CLI we're on. Um we also have all these shared skills that we kind of share across the team and allow people to create their own channels of skills and then when skills

38:53 are kind of good enough or they are um sort of central to the process they'll get merged into main. So I can just show here we have um if we go 10x skills status you can oh 10x

39:12 it's a typo 10x skills status um you can see that uh we have all these skills that correspond to every single one of our artifacts. So if you're creating one of these structured artifacts, the agent

39:26 knows exactly that structure and and how to um how to build that artifact in the in the right way. You can also see if we do um 10x context and then we make the mode for the agent, it's going to list

39:41 out much more detailed information for for the agent to see and it will actually list out every single one of these artifacts. So it knows exactly where to find all the information about

39:54 the epics, the architecture, the different conventions, um, and a log of all the recent stuff that's happened in the repo. So that would include, um, it's basically always writing back to

40:06 this log after it's doing any significant work. And we have all of this running on a start hook. So every um every agent session that anyone at 10x is running, we're basically we'll

40:20 call it like benevolent prompt injection, whenever the agent wakes up, it's getting this kind of packet of context that's telling it all about its um workspace and every artifact that it

40:32 has access to. Um maybe I'll pause there and and >> yeah, I have so many notes. >> Yeah. >> Okay. So I want to first think when you

40:41 open this up like I noticed the separate uh repos right. I wanted to ask you guys is it true for you that like over time the context repo is actually becoming more valuable than the code repo.

40:52 >> I certainly think so. I think like the percentage of time that you spend on like engineering the markdown should be higher than the percentage of time that you spend on actually executing the

41:03 code. I can spend like four hours just building really good plans and architecture documents and kind of setting up this whole project management repo and then I can just say like

41:12 slashgoal slash10xprocess slashexecute project spec and I can just kind of put my computer in the corner and go to bed and wake up with exactly uh kind of what I'd built. And then the

41:24 beautiful thing is that with this system it's all all of the tickets have been moved to the right column in linear. everything has been written back appropriately to the to the repo, you

41:35 have just like full um trail of what it did. And so I I definitely think that the the information and kind of the meta harness and all the markdown is is a lot more important than than the code.

41:47 >> Would you agree, Alex? >> Yes, I totally agree. I mean, if I think about just the highest level as like an as a non-engineer, what is valuable here? It's just that like there's just a

41:58 set of norms and customs that keep an agent on the rails and then the whole issue that people navigate right when building software is session to session memory or context. And so I think this

42:11 bu benevolent prompt injection that Dan talks about is kind of our solve for making sure the exact right information is provided at the exact right time without lapses so agents can work on

42:22 longunning tasks. Yeah, my question was like uh even deeper like if if imagine like if someone you know hacked your system and like they could only steal one repo because like personally you

42:31 know my code is like okay it's already being written by agents right like I I don't care anybody could replicate it but but all the other stuff like the core business insights

42:40 >> yeah the process the ideas the customer insights like the the marketing strategies I have the reason I think it will work like those are the things that I still value because you know they're

42:49 human >> well I think in general and Den I'd be curious your thoughts as well. It's like just think about like traditional markets like economic markets in terms

42:57 of supply demand. It's like there's going to be the most demand for where there's the least supply and there's more supply of code than ever before. There's way more scarcity of I would say

43:12 unique process or business customs or internal knowledge that creates alpha and so of course that is going to over the long term be more valuable. I would definitely agree with that. I would also

43:23 think of it through the lens of like if you had to reverse engineer someone's product, would you rather just have their entire codebase or would you rather have the the project management

43:33 repo? And I think first of all, the codebase becomes really hard for the agent to digest even if the codebase is like the ultimate source of truth of what exists in the project today. it's

43:44 sort of devoid of all the meaning and structure of like why this feature exists and how it relates to everything else other than the like direct code links. And so I actually think you could

43:56 take a project management repo and probably reverse engineer or even like build a a better product because it contains all that um intent and conventions and the architecture and the

44:08 way that you run the system and and all that kind of like built in um which is a total like flipping things on its head from the way that engineering used to be.

44:19 >> I I was going to say one thing and David you may find this interesting. I don't know if Dan there's anything to show for this, but one thing I've like as I learned more about our process that I

44:28 always found interesting is this idea of how do you make more and more things outside of like the pure code verifiable and I feel like the whole like validator step you have is just like a really

44:38 interesting piece of this pie. >> Yeah. So I can I can go ahead and say 10x validate um see what that'll give me. And so

44:49 immediately we're we're finding basically these warnings that it's going through all of the artifacts that we have in in this project management repository. And then there is a a really

45:00 long rule set that kind of aligns with our just SOPs for how we like to write software and all the steps in the process of taking um customer requirements or a problem that we're

45:13 trying to solve all the way to like a landed PR or a landed product. And so in this exact in this example basically has um the checkpoint status is drifted in this project spec. And so what that

45:30 means is that the authored status where it says complete as the status of the checkpoint is disagreeing with the derived status of in review and that derived status is basically computed

45:42 from the rules that we've set. And so in this case it might mean that um the status of that project spec says it's complete but the project spec actually has tickets in inside of it that are not

45:53 yet complete or the tickets are in review. And so you can't resolve the full project to completed if not all the uh specs are if not all the tickets are completed. And so you can just kind of

46:05 extrapolate from there and multiply that rule out and we have kind of hundreds of these types of rules. So it's kind of like linting for your SDLC or your entire engineering process and not just

46:18 linting for the code and and this it also takes advantage of this property of agents which is that the agent can use the CLI and so the agent can constantly be figuring out what's the most

46:30 important thing to do next what is drifted and how can I how can the agent fix it and so it's this uh very nice like self-improving loop. Yeah, I think there's tremendous alpha in actually

46:40 having agents hold humans accountable, right? Like if you say like these are the priorities of the company long term. These are the biggest projects for this week and then like somebody gets

46:49 distracted by like a cool idea from Twitter. It's like listen this is this is not on our road map. This is not you know top five priority. Why are you working on this? I think people need to

46:58 be like way more comfortable giving the agents the things they're great at. Right? Because with each generation of models the agents are actually better than humans at more things. Everybody

47:08 understands is already like digesting lots of content. You can take a whole book, 10 books and like read it instantly, right? But like consistency, predictability, all these things humans

47:18 are really bad at. Humans are kind of random. They're creative. You know, they like to switch from task to task. But agents I think will be used more like to hold us accountable. Like for me like I

47:28 have many skills where like build as much as possible yourself and then once you're blocked by something like me walk me step by step and literally I just send it screenshots and it helps me do

47:36 some setting in superbase or something where like normally you know I would just probably take way longer and the agent would would be like writing a little bit of code then it's blocked by

47:46 me then you know speak more about this you know not just building for humans but building for agents because you mentioned that earlier and I think this is where the world is headed. Yeah, I

47:53 think I think you kind of have to just think so creatively about the way that software should be written now. It's kind of like everything you ever learned about building software is is kind of

48:03 thrown on its head. And I think we usually I I hear this like negative framing a lot of times which is like the agent is deficient at coding or or work uh working with the codebase because

48:15 every session um starts from it starts from zero knowledge of the codebase and you have to kind of like build up its context and so only the human can like maintain this coherence across many

48:27 different sessions. And I think if we think about it from the perspective of like the things you said, consistency, um, structure, how can we like make use of all the advantages of agents and then

48:39 just solve those deficiencies? We can get a very interesting result. And so one thing we do is basically um we think of every session as uh the agent should start as like a senior engineer on the

48:53 project. It should immediately know what's going on. >> So does the hook. >> Yeah. So the the hook and then if we I'll just uh change this to low so it

49:01 goes a little faster. If we use this uh 10x process skill, so it knows what to do and I just say like what should I work on next, it kind of immediately see it runs that hook with with all the

49:15 context that you saw earlier and then it's going to use some 10x process commands and pull in kind of the exact things to work on next. and it can also pull in the the architecture. And so it

49:27 just solves very naturally that problem of like the agent doesn't know what it's doing when you start a fresh session. Now, not only does it know what it's doing, but it's following this extremely

49:38 consistent um and rule-based process. So, we can have um kind of like level up the humans that are working this way as well. >> Did you guys open source any of these?

49:47 >> It's kind of uh an open discussion uh of what parts of this we we will open source or not. So um maybe maybe in the near future we we will do something like that.

49:57 >> I would think about it. You don't have to start with everything but like you know when I released my skills repo it quickly became my number one GitHub repo and there are way more popular skills

50:06 repo than mine. So I think this is a very good strategy to get some motion. >> Yeah definitely >> appreciate guys. Thank you for your time. Where should people go?

50:14 >> If people want to learn more about what we're doing at 10x10x.co co. Um, we are hiring exactly one ton of people. Uh, cracked engineers, AI strategists, full stack creators, you

50:27 can go to 10x.co/careers. And, um, if you want to just follow along with the content we're putting out, follow me on X at business barista because not only am I creating content,

50:37 but I'm amplifying all the content by folks like Dan, CJ, and other people in the company. So, you'll see kind of everything from the what we call WWE for nerds, uh, from my account.

50:47 >> All right. Awesome. I'm going to put all those links below the video. So again guys, thank you for your time and have a great day. >> Thanks so much.

50:53 >> Thanks so much. >> All right, so here are the results from the scrape. Uh as you can see, it found the AI companies valued at over $1 billion and it found who the CTO is and

51:02 it started looking for the two forms of contact for the CTO's. And again, this is just a single use case. You can imagine hundreds of different things you can do with deep API. In fact, if you go

51:11 to the docs, you can look at API reference and see all of these different endpoints. And again, all of these are available through a single API key. No setup required, no configuration or

51:22 authentication. You get all of this right off the box. And now, as I mentioned, we're opening up deep API to the public. And we're giving every new user $1 of free credits. So, just go to

51:32 deep aapi.co and give it a shot yourself. See how it works. Try scraping something. Try accessing some difficult website, doing a deep research, looking for more clients, whatever you want. You

51:43 know, try it yourself. Just go to deepappi.co. It's completely free. We give you $1 to play around and test it yourself. Or it's also going to be the second link below the video.