will ai agents replace devs or just 10x them w/ peter bell

full transcript

auto-generated from the audio, so expect a few typos. the audio is the source of truth.

All righty, good morning. Internet happy Thursday. Excited for this episode. You know, again, I need I need a shout out. Boardy, you know, again, for connecting us. Actually, a few of the guests recently have been with Boardy. And I Peter, I got paid $2 yesterday for tweeting about Boardy. So there you go. Who knows what you'll get paid for mentioning him on a live stream. Yeah, don't don't don't send, you know, don't don't spend too. I won't spend it all in one place. I thought you'd die Coke this morning. Yeah, yeah, yeah. So, so, but again, shout out, Boardy, you know, I think and we're going to talk about today.

I feel like this conversation. So whenever I get somebody like Peter on the stream who can yap, I feel like conversations will kind of go all over the place. But I think I think the biggest thing is right now is like, if you're not experimenting with some sort of AI in yields, I think you're just heavily missing out, right? I mean, like, you know, people will be like, well, Taylor, you on boardy to find a job. I'm not looking at boarding to find a job. I'm looking at boarding to see what other companies are doing. Because essentially it's kind of in in in the same area.

And so, you know, for me is is I'm I've signed up to every single AI platform I can that I've seen to see what they're doing. I, you know, what's my experience with it? How can I speak to it with devs? I mean, you know, I told Peter I was I was on a call with a developer this week who got laid off after just a few weeks. And I told him about all these sites and I hooked him up with, you know, a, you know, a free billion tokens from Muse. And it's like, listen, I mean, you know, you know, and the guy the developer on a call like, yeah, I didn't even think about it.

If I can use a gentic workflows for my job search, then I can put that on the resume. I like, yeah, literally, like, like, like, you got to figure out how to do this. So anyways, building open sort, building open source to help other people to refine their job search with a research pipeline. You can just use a lightweight supervisor pattern in the skills rather than having to go to a factory and deterministic orchestration and durable execution. And all that stuff. There are lightweight ways that you can build tools, add rigor to them. And I mean, we'll talk about it more later, but the way that you get a job now is by showing proof of work because it's close to free.

If you can find 20 bucks a month, somehow, you know, extra couple hours at McDonald's, whatever you need to do, you can afford the tokens to show what you can do. Well, that was interesting. And again, we haven't even kicked off the show yet. But again, I knew this was going to happen because I just wanted to get all like all this. I thought thought value out of Peter, but, you know, as developers said something interesting to me, he was like, I don't have a lot of money right now because I'm in between jobs. Yeah. So I can't get out here and code a bunch of stuff.

Like, you know, because I can't, I don't have money to spend on like, you know, the Claude Max or the cursor pro, whatever, right? Yeah. And it's like, I was like, yeah, but like you have an instinct. You have a muse where you get a ton of tokens and the free plan for muse is actually very generous. Like I have, I haven't maxed it out yet. And I'm running a lot on muse right now. And so like you got to get creative. And but I thought that was an interesting kind of thought that people on the jobs just don't have money. It can't get the subscription.

But I think there's ways around doing that. There are it's important to both acknowledge life isn't fair and to say wherever you are is where you start from and what you've got to work on. Like I was, I ran the instructional team at Flatiron School, which was a bootcamp, General Assembly before that. I know lots of amazing human beings who had the capacity to be great software engineers, many of who became great and successful software engineers, many of them, you know, what you've got an elderly mom, you've got two family members to take care of because they're like parental. So you don't have the time or the money like there's this kind of like, oh, you know, I figured it out all on my own.

Yeah, but you know, somebody else paid for college and you could spend evening some weekends coding on a max plan your parents paid for. So it's not fair. It's not free, but that doesn't mean you can't dig out from wherever you're starting from. Yep, absolutely. Absolutely. All right. So first off, good morning, everybody. Good morning. Excited for this episode. So if you are tuning in for the first time or listening to this in the future for the first time, I've been doing this podcast series now for six years. And, and you know, this really started out. I got on LinkedIn live when it was in beta super early, kind of started talking during COVID when people were obviously very concerned about where the market was heading, just talk about some basic blocking and tackling things you can do for your job search.

And then it's kind of really evolved over the last six years to bringing people on the show. And really now it's evolved again. And I just want to talk about AI and, and, and I mean, and really, you know, I think there's a lot of eye rolls there, but I think it's the state of the internet. I mean, I think I think this is where we're at. Like it's not, Hey, let's bring you on to talk about the latest JavaScript framework. Hey, what are you doing with, with, you know, a gentic AI now? And, and then, you know, my, my whole kind of thesis is, is giving people inside about your career.

And what's happening, which Peter's going to shed light on today. So again, excited to have Peter on the show. Please connect with him. Highly recommend you connect with him. Send him a message. And again, as always, we are streaming across all the channels. And if you are not in our tort discord yet, please do so. And again, I always got to plug torque. If you're on the job search, I think you should sign up. I think it's just another tool in your tool belt. Again, we are connected with Ronstad Digital, the largest staffing company in the world. We get hundreds of jobs across the globe.

We have an AI matching algorithm that matches the jobs in our ecosystem. So again, you know, we, we are in a silver bullet. And I think one thing we're going to talk about today is I don't think there's any silver bullets right now in terms of the job search. But, you know, again, we'd love to have you sign up. We'd love to have you part of our community. And again, we have a very active discord. We do 20 events a month. Shout out, Roxie. She's on the stream. She's my colleague. She crushes it. And so anyways, yeah, excited. All right, let's go and dive in here.

Let's get to people what they want first, man. Who are you? What do you do? And we'll go over there. So, you know, I used to write software for a living. Then I wanted more impact. I became a CTO. I ran engineering at General Assembly. Bunch of other startups you haven't heard of, maybe 15, 16 years ago. Then I'm like, wow, you know, CTOs, this is really hard. And what I wanted to do was learn from other CTOs, how to get better. So I started. I co-founded the New York CTO school. I created CTO connection, a series of CTO communities for directors, VPCTOs. We'd run single day single track conferences, New York, San Francisco, and around the world.

And then these days, what I do is I run gather.dev. We're just relaunching. So everything's going to look different tomorrow morning, I promise. I keep getting pushed back by consulting work. But basically, what I want to do is build the contacts graph for agentic engineering. So turns out we were all asking more or less the same questions. There's the technological questions. What are self-worth factories? How do we build fields, head skills libraries? How do you do deterministic validations? What kind of DevOps policies, procedures and systems? Do you need canary rollout, you know, feature flagging, all these things? And then there's the engineering leadership culture, stakeholder expectations, creating an environment where you can learn what's the future of roles?

How do you hire for people in the future? So what I'm doing is building out a taxonomy, researching third party data, and then saying to directors, VPC, he goes, hey, you want to get in here? Give us your benchmarks. We'll give you someone else's. And then instead of going in the slide channel and saying, hey, do you have an answer about X, you literally gather.ask. And whether it's you or your agent, you put in the ask with you, your context for your company and your question structured. And then it gives you relevant feedback from similarly sized, similarly mature companies. I was still building out the knowledge graph, but that's my passion.

I also am writing the Arilie book for scaling AI adoption in engineering. Hey, when that's out, we need to get a discount code and get a community. We can actually do better than that. I'll find you the link. Antithosis is giving it away for free to anybody who wants a digital copy. So I will get you a link. Yeah, I want to get a free copy link afterwards and get it on get it on our community hub. And then the final thing I'm doing, because this is a transition. I still see some CTO saying, I don't know, like I I managed mobile and I managed, like, you know, on prem to cloud migration.

I'm just going to hire a VP to figure this stuff out. You know, CTO running 500,000 person or I might do it. If you're not using AI, if you're not shipping code, maybe not production code. If you're not shipping code daily, you have no idea what's going on. So I spent the last six months abusing 12 max plans across anthropic open AI and cursor. I was at API rates just for my personal use. I've been spending about 50 to 60,000 bucks a month. And I went from two or three GitHub commits a year because, you know, pointy head boss too busy to code. So I've merged 23,000 PR since January on my own.

Wow. Crazy, crazy, crazy, crazy. All right. So, you know, I got a course, my AI to research Peter last night and I come up with questions funny how that works. Also shout out, Pally. I've been obviously I try all of them. I have faux on here. I have Pally. I have instinct. I have muse. I have Grockbot. I have all of them, right? I'm trying all of them. But I want to give a shout to Pally. I really like what they do. They called and made an appointment for me yesterday for my corgi to get groomed. And that was awesome. I just said, Hey, I need you to call this place.

I need you to go and book it. I don't have time. Check my calendar. It found a calendar. It found it when I got back from the beach with the family. It called them. It booked it. I mean, it just. So again, if I do have subscriptions, I have unlimited subscriptions for Pally Pro. It's a free month of it. So if you do want that, please let me know. But anyway, so Pally research Peter. So Peter, we got a lot of questions today. You got a lot of questions and let's do it. I want to go with what the title is around the 10 X, right?

Will AI agents replace devs or just 10 X them? And so I want to get this question here around org restructuring and 10 X engineer myths. So a lot of execs brought into the narrative that they can hire 70% fewer devs, want, want, want, and just to make them three X faster. You focus on heavily on org design. Are engineering leaders actually shrinking headcounts, permanently? Or are the shapes of engineering teams evolving into different roles entirely? So great question that should give us a 45 minutes. So many things. Firstly, it's important to realize the future is here. It's not evenly distributed, right? So I speak with Farhan, who's head of engineering at Shopify.

I speak with Rob Widoff, who's a CTO at Coinbase. Did I spoke with Mike Beauford, who's used to be CTO at Grant Greenhouse. He's now a member of technical staff at anthropic Ryan Luppa Polo, wrote the harness engineering piece at OpenAI. Yeah. And those people who do an AI like for real. The vast majority I was speaking, I'm talked to a CTO who runs a team of over a thousand people over the summer. And he's like, I don't like we're doing AI. I don't get the point. I've only got so many experiments. I can run with the data. I've got a thousand engineers to build it.

What would AI do for me? So there's still a distribution of opinions. And the good news is your exposure determines how crucial this is. You know what? You want to change imps. AI is not suddenly going to make muscle for humans without them. I haven't to go to Gold's Gym or whatever. So if you're in the business of running gyms, sure, you can get some operational efficiency by using AI. You don't have to. If you're in the business of, I don't know, building online websites, you're probably going to see some real AI competition sometime soon, if not already. So you have to be a lot closer to like, we always talk about technology, adoption, lifecycle curve, like innovators, early adopters, early and late majority and laggards.

Where you sit on that should be a corporate decision based upon your risk and your ambition. You know what? You could run a chain of gyms, but you want to run every retail store in the world, in which case, maybe you do need AI to get into groceries and 40 other things. So first thing is to understand it's not evenly distributed. Most of what we're going to see is going to be in 27 and 28 in bigger companies. Because a lot of CTOs I speak to are like, well, you know, we ran experiments and 25 was like, we tried it. It's like it was fine.

November 25, it started to be good. 26, they ran experiments. And now they're doing headcount planning for 27. So the context first is it's very distributed. This is going to take years, not days. I agree. Yeah, I agree. I mean, even even with our subset, with, with my team, right? I mean, I think it's, I think it's on the leader to pick what tools make sense. Um, but then also slowly adopting it, right? I, I think the biggest thing I'm seeing is that companies are doing like, Hey, we're using this. Take our online training and then, you know, congratulations, but I, exactly. For me, it's like that, that like adoption goes in on a very individual level that is going to take a lot of time to scale.

So there, there are three groups. There are the, the five reasons that there's like the inventories. There are like the settlers, it's majority. And then there are the resistors, the people who just never going to do it. Uh, to be really frank, the way I see that plane out over two years and you can move from any category to another. The fire breathers are going to do great. They have career security because they are the people learning to make these machines sing. They are the people when, and I have no affiliation with this committee, you know, a pick up couple months ago, they fired 20% of their team, but also said, we'd pay an IC a million dollars if they got enough robots to deliver enough value.

I think that's the MBA application. Uh, Gergie, the pragmatic engineer, great publication. Uh, he talks about this trimodal distribution. I think we're going to see a lot of weight, a lot more weight in that third level. I was speaking to somebody who was in that third, but you know, there's like, they're going to crap you, but it pays the bills. There's the solid big tech and then there's that absolute killer salaries where ICs are making, you know, 700, 800, 1.2, 1.4 million dollars a year, uh, total comp. And I think we're going to see a substantial growth in the number of ICs there. Uh, so you, if you're an early adopter, great.

If you're spending evenings and weekends, if you have that luxury, that, that privilege to be able to spend your spare time and maybe a little spare money back in the day following Steve Yege these day, you know, Justin McCarthy, looking at what the labs are doing, looking at the trends, Simon Willison, all the players, great. Uh, most companies have three to 15% of those people. They're really lucky if it's 15, most often at enterprises, three to five percent. You've got 60 or 70% they just like to just tell me what to do. So what you do is you get that as an engineering leader, your goal is to get those fire breathers to pave the cow parts to make it easier.

Start. We'll, we'll just show you how to build a scale. Now we'll build a shared skills library. Now we'll improve the quality of the shared skills, then we'll trim the number of them and show you how to leverage the good ones. Then we should supervise the patents for subagents. And before you know it, 18 months from now, you've got a software factory running and people using it. Absolutely. Absolutely. Um, let me. Uh, so I asked, I brought on Nader Dabbitt, whose head of growth at cognition. And I asked him this question. I was like, are we, are we going to be in a place in, let's say 2030 where it's just going to be like a very, very lean team and a bunch of agents just running an orchestration of agents or do you, or do you think we're never going to get there across the board?

We're going to get that because you have to, the first thing is you have to generate good code. So first thing you do is you generate code and it sucks. And the best analogy I've heard for this is playing the guitar. Do you play guitar at all? I play drums. Great. So let's say I gave you a guitar. You strum the thing a couple of times. You're like, this thing's broke. It sounds like crap. Right. Well, no, you got, you got to get a little more practicing, right? You got to, you got to put your 10,000 hours, whatever it is. So it's not the guitar.

It's you. And that's pretty much where we're at with agents now. The models are already good enough to write better code than pretty much any human. We're just doing a very bad job of actually telling them, we will say, Oh, build a site that does this. And then we're like, Oh, but it's not secure. Well, we're in the prompt. Did you ask for it to be secure or well architected or maintainable or have a clear domain language or, or thog analogy of API endpoints or wrap, you know, rigorous typing to avoid, you know, type errors in your inputs and outputs. You've got to ask for what you want.

We're getting better at doing that with harnesses, context, doors and systems. That's the work. So firstly, generate code that doesn't suck. We're getting there. Secondly, I've tried all the things you need to have deterministic verification. Like the people I speak like the Jeffrey Huntley, the guy who wrote the Ralph Loop thing. And I've run a lot of Apollo who was at OpenAI now at Google. These guys, they're literally, I mean, Jeffrey was talking about, I'm abusing the compiler, like I'm getting a compiler to emit an AST and abstract syntax pre and then we're deterministic checks against that. The future is going to look much more like people doing it like these four more methods.

Going to look for correctness, much more like formal verification. But these are problems we're going to solve. And then the third problem is how do we come up with enough experiments and ideas to figure out what we should build. You can solve all these problems. Yes, very small, very lean teams with huge numbers of agents that does not necessarily mean there will be less employment for engineers. It just means that for any given company, they either need to expand their ambition, improve the quality. We're going to run 20 different experiments and then only ship the one that really wins or they might have smaller teams.

And I'm hoping that the balance across the industry will be Jevons paradox. NGA can deliver a lot more value than they ever could. So they're worth hiring. But there's going to be, I have so much software that's an N of one. Lots of small companies are going to have a lot more software. There's going to be FDE roles for deployed engineers across automating finance, accounting, customer service, operations. Going to be plenty for AI-filled engineers to do for the foreseeable future. Dude, there's so many tangent questions I want to ask you with this. Let's talk about and again selfishly, I'm just asking you about these trends so that I can speak to them.

And obviously my circle for deployment engineers. Let's talk about that. What that and again, of course, I created these questions and we're not using them yet. But we got one of them. We got one of them. What what is a for deployment engineer? You know, I know some people want to get into that more. Is that something you can actually get into officially? I still feel like it's like you kind of got to do the job before you get the job. Let's let's let's unpack the for deployment engineer. Yeah, so I think that as far as I'm aware, the forward deployed engineer, the the the really the starting point for this was Palantir, who are well ahead of the game with his stuff.

They were already in the business of basically business processary engineer and how can we take this stuff? A bunch of inefficient stuff, re-envision the outcomes we need, the checks we need, the information we need to capture and share and synthesize and rebuild those processes. AI obviously made that a lot faster and easier. And the the heart of their learning was that you need to build these rigorous ontologies. You need to figure out basic build a context graph of the things, the people, the companies, the roles, the tasks, the projects. And that's what allows you to bring enough rigor that you can then run queries against that graph to give the right context to get the agents to be less done.

Problem is all the stuff that needs to go in, like what should the structure of that graph be? What's the answer? What's this edge like this connection between this person and this company and this contract with this claim we're trying to process? All of that's usually some of it's in the database somewhere, a bunch of it's in Excel spreadsheets and the rest is in Mary's mind. And every time Mary comes into work, she knows exactly what combination of spreadsheets and databases and other things to click on to figure it out. The job of a forward deployed engineer is to sit down with Mary and say, so Mary, looks like you've got a bunch of kind of want work here.

My job is to try to make your job suck a little less so you can actually focus on the higher level. And you've got to be realistic here. They're going to be scared because like the tacit knowledge is the one thing that makes them employable, right? You know, are they going to lose their job if you implement that? It was back when engineers in the States were trading their Indian replacements who then flew back to India and got fired. So it's hard. Some of these people will have higher level roles. Some of these people, there'll be less of them. So it's a difficult environment, but a forward deployed engineer is somebody who goes down and sits with Mary and tries to help her so that her job sucks less.

And does that by asking all the questions, looking over her shoulder and then running lots of experiments, like, maybe we should like just create a little graph here and maybe we could like ingest this straight from an API and maybe instead of a spreadsheet, we could give you an interface that allow you to type that in, but there would be a little more structured. So we knew what every field was and it would go into the right place in the next step of the process. So I have a heartache. So you're saying a four deployment engineers, just a business analyst who can code. Yes, absolutely.

I would not, but I would not say just. I would say, you know, it's like, you know, they just an MBA player who won a couple titles, that is one of the most important skills you can have. If you're a business analyst, if you can actually speak to humans, not be miserable, not make them miserable and deliver outcomes, that deliver value to the company and hopefully elevates that person's role. I mean, that that is I mean, to me, that's live in the dream. It's an amazing value. So I spoke out in Raleigh and I'm actually trying to ask you these questions because I just want to send this episode to this one person to listen to because I still remember the email went back and forth with him on.

So would you say like, so what do you think could be a person? So if if a person was in a role and they're listening to this now, they're like, Oh, that kind of sounds like what I'm doing. What are what are some specific things a person would need in their job to be to be qualified as a four deployment engineer, even if they're not classified as that on the actual title? So it's a good question. I have not personally hired for FDEs, so I'm just going to give you my best take. I've got that's fine. That's fine. Some I did actually hire somebody who was running solution engineering, you know, one of the AI teams.

One of the things that is and this broadly goes into what we're looking for. There's no question that it's very much like what used to be a solution engineering role, customer success engineering role. You are engaging with humans, making them happy and quickly shipping value with some rigor. And finally, with a volunteer model, if you're doing it right, you are providing product feedback to the platform team to improve the underlying infrastructure so you can solve the next business problem a little quicker because we need a shim that does this or we need an integration with Workday or whatever it might be. Yeah, absolutely. So the things you want to look for is proof that you I don't know that this is so much a resume matching thing as a capability is matching.

And FDE, it's basically you need to prove you can speak to humans and you can prototype and deliver value quickly, but that you have sufficient rigor in your thinking that you can think in systems and propose intelligent improvements to the underlying platform. Yeah, I love that. So so I want to go back to this is popped into my brain, my ADHD brain. I was on obviously Open AI did their dev day yesterday and and shout out Pauline. She's a friend of the Torque community. She emcee the stage. She just got a job at Open AI from Versailles. Shout out Pauline. She met Sam Altman yesterday and got a selfie.

I thought that was really cool. Um, you know, Sam, Sam did an interview with the CEO, uh, Dan of every, every, yep. And, um, he made a comment that you just made a comment of, and I want you to unpack this a little bit more. Sam made a comment about how because of AI, we're able to launch 30 ideas at one time and see which one sticks and you kind of just made a comment about that. So I'm curious if you want to unpack that a little bit. Absolutely. So just speaking with Jeff, God, Elf, I'm, uh, will be dropping the interview as soon as I get my robots.

Well enough. You tell the robot. I wife calls them robots. He goes, are you talking to me or your robot? I go, which robot? I was like the instinct one. It even was mine have names. So I, it turns out that humans are actually quite there. There are interface patterns that work well for humans. And one of them, I've managed teams of humans in the past. So actually having a lease who's my editor in chief Sloan, who's my head of engineering Morgan, runs the, as the C.O.R. We got Nate, who runs the sponsorship basically is the head of sales. Turns out that that's a useful context management strategy for the human.

It actually works well. So yeah, you can build 30 things. And speaking with Jeff, the question is, should you, and the most important, one of the most important skills I think you can build as a developer. Now there's this two skills. Three skills. First, speaking to humans. Long as there's other humans around in the company, if they liked you more, you'll probably be around a little bit longer. So it's pretty good to get good at speaking to humans if you can. And I know that's not easy for all of us. The second part is you need to get good at AI and we can dig into all the details, but you have to be sure and prove for work and get the experience you can with the tokens and the time you can reasonably find and afford in your life.

And then I think the third part of it is you need to get really good at product. There will be no job in 2030 for somebody to pull linear tickets and ship code that meets the spec. That job is gone or is gone away. It's gone already. It's just, you know, it's like, you know, somebody's dead. They just don't realize that that is we're not Java programmers anymore. Or I moved from code fusion to Ruby to a bunch of other things. And it's like, Oh my God, are you late? Like I love code fusion. I love the people I met in code fusion. I'm not a code fusion developer.

I'm somebody who solves business problems and the code just needs to end up in assembly on the chips and work. And there's different tools for different ways of doing it. So our job is not to write code. Our job is to solve problems. Broadly, I think engineering is going to bifurcade into two broad roles. And there will still be lots of specialists who bring value in a different way. Broadly, you're either platform devx or you are stream aligned or product engineer. Job of a product engineer is to deliver value to customers, to come up with hypotheses, experiments to create like a lean product canvas, figure out what the pain points is, what we think the value would be, what would change in the world if we offered that and what would be the simplest experiment we could run to see whether there is demand and technical capability to deliver that.

So that's the product engineer. And that's a huge part of it. And then the other part of it is the platform devx engineer, which is building the harnesses, the systems, the evaluators, the loops, the tooling to allow the product engineers to be more efficient. So the good message you can build 30 things. The bad news is you've now got to bring a lot more rigor about which 30 things you choose to build. And it's way better to start from an audience with a pain point, with a hypothesis that a change in your system would provide a change in their behavior to provide a valued reward for them and you.

And then one of the things you generate first, go generate 30 ways of getting that change in behavior. And then maybe lots of ways of experimenting, whether you do user groups, whether you do focus, whether you set one on one with people, or whether you do canary rollouts for a subsidy or population, use and feature flags. But you've got to use that to figure out to get in the room with people and see if it makes the change. So good, Peter. You know, again, like I so I have a blog going to drop next week. I was in Dallas for 36 hours at a committer code conference and I went golfing.

I'm a big golf guy. I don't golf as much as I want to because of three kids, but I'm a big golf guy. And and I caught up with my buddy Ben Pyle. Ben is a Ben was a former CTO. Now he he really is kind of moved into a consultant role. And and specifically he's full A.I. Pilled at this point and he and I and I have it on the video. But what you just said, I even see. Yeah. And if I see it as a non developer, other developers need to catch on, right? Like, like I consider myself the normie, right? Like I'm in kind of this weird kind of intersection of like recruiter, but not recruiter or super embedded engineering community, but not a coder.

Like I'm kind of this weird intersection of like this normie guy. And I'm even seeing the transition to where this is going. And I think to me, it's like, like, like I think industry, industry used to not matter. Why? Because you could code whatever for the right now. I feel like industry is somebody's greatest weapon. My industry is making an AI platform feel more human. Like that's, that's my lane, right? And so for me, it's like, if you're super ingrained in the mortgage industry and you're using AI or you're super ingrained in the healthcare industry, that is your mode now. And I think you should double triple down on that.

It's going to, and it already started to happen. I feel like there was overhiring during the pandemic when we thought that everyone would always be online for everything. Then we found out they weren't. And so there was this pullback and there's a lot of senior engineering leaders. I know who were like the market got so weird for jobs, especially for two, three years after like, you know, 2022 or so, where it was like, well, you, you have eight years experience at a series B venture back health tech company, but it's not the right subset of health tech. So not going to hire you as a CTL.

Right. And I think that it's great that you are helping connect people with jobs. Because I think one of the biggest challenges we're going to face, something that I'm going to be launching on gather is gather.next. Hopefully at the weekend, which is how do you figure out where you want to go next? What's the future impact you want to have, whether you're in between jobs, or you're just thinking about it? And then how can we build a really good, like Spotify did for music? We think you're going to like this song. How can we pick the companies we think you're going to like? You're going to succeed well and because I want good and bad devs, there are, but they're also great devs who are horrible in a certain environment.

Maybe they need more guidelines. Maybe they need more freedom. Maybe they need other tooling. Maybe they need a different pace or a different problem space. All of those are going to become increasingly important. But as you said, the more you know about your industry, the more differentiated you will be in roles for that industry. And now we just have a much more precise matching problem. Because now there's only seven engineers who'd be great for that role in the whole of the US. We've got to either find one of them or persuade one of them to join. I love it. I love it. Let me ask you this.

This kind of goes to one of our questions here around tactical career advice. Right. So I ask Nader this. I asked Aaron Francis this who's head of VP and marketing for Laravelle. How like people are like, all right, I do the muse. I do the personal agents. Right. How do I get started with AI? What do I do? What are what are some? And in Peter's expertise, if I was a developer, let's say I'm a five year developer right now, right? I'm a level developer, three to five year developer. And I'm listening to this or watching this now in the future. And I'm like, man, Peter and Taylor are just spitting some heat right now.

This is all great and all. But what are we going to? But so how do I get started? So there's a few principles. I love the muse instinct thing and I may well get that appealed, but especially as developers, if your entire interfaces like this kind of amorphous thing that just does it for you, you might be getting some value from what it does, but you're getting no value from learning or limited value from learning. So the first thing you got to do is you got to own your own learning. Simplest thing you do, Claude code, code X, cursor, whatever one you use, get the plan you can recently afford, start cheap, you can upgrade later if you can afford it.

That's right. I had somebody ask me this a while back to like, how do I know which plan? I said, start free. Yep. If you max it out consistently, go to the next one. And if you start maxing out, go to that. Like that's what I did. Exactly. Unfortunately, if you're not careful and you have the resources, you end up with 12 plans, but it was $12, $200 a month plan. It was a thing, but it was worth it. So wherever you're on that first thing you do, check out your start with all code. Oh, code X, if you prefer, doesn't matter. Basically tell it to create a repo for your repository on GitHub is things changed with GitHub probably, but it's fine to have it for now.

And then just say, OK, at the end of every session, I want you, could you make a little skill called a remember stuff? And then any time that I say, remember this, and then at the end of each session, this is very naive. I'm just going to ask you to run the skill called remember stuff and it will remember stuff that it thinks useful. So now what you have is a repository that's learning a little bit about you. Real quick, real quick, talk about a skill. I haven't, I'm about to dive into this stuff on my sessions and I'm doing personal streams, but talk about a skill real quick for people who may not know what a skill is.

So basically a skill is a combination of prompts, words and optionally code that allows an agent to repeatably do a thing. So it's like you sit down with an intern. Let me show you how we make coffee here. This is where the stuff is and this is where the water is and you've got to remember to fill this and the beans are over here, but don't use these beans, use these ones first. That's the scale. It's the steps it needs to go through to do the thing, whether it's remember stuff at the end of a session, make a commit, write good react code or fill out your TPS reports or your JIRA tickets or whatever.

So you start with that. You can go a very long way down that rabbit hole, but what you want to start doing is learning what you need to do to direct the robots to efficiently achieve your outcomes and it's the same as the guitar. If they suck, it's you, not them. It's not that the robots can't do it. They can. Like I have completely failed at getting robots to write pros, like, write, you know, articles and books, the Riley book. I still don't like the stuff they write. The reason for that is even though I spent months on it, I've still not figured it out.

There is something that I need to do, a piece of rigor, I need to bring a tweet to the prompt and one day that will work. I mean, I'll be doing everything. I assume it is and that'll get you. But then look for the easy wins. I'm coding. They're really good at and you can make them better and just get it to start to understand more about your context, your environment and your problems. It saves it all to the repository and then it's a rabbit hole from there. I look into supervisor patterns where you tell an agent to go tell other agents to do stuff, then software factories, dreaming systems, context graphs.

It'll take the rest of your life, but every step will make you more differentiated because then when you go into a conversation at a company, they're not like, oh, you still know how to do RAS. They'll be like, dude, you can talk to us about tuning our verifications and validations and set in our loop budgets. That's what everyone wants to hire and the way you get that job. It's 20 bucks a month or it's free. The way you get that job is by proving the agency by finding one night and doing it. So my yeah, and Aaron, Aaron Francis said this in our stream. He was like, I think people still get frustrated that it doesn't work and they just walk away from it.

And the reality is like you have to keep at it. Like you have to keep at it. Go ahead. Yeah, you're exactly right. Right. So my hypothesis and it's wrong, but I pretend there is nothing that I can do that the agents can't do better. Yeah. What that does is it allows me to keep trying and then keep experimenting. And there's been a whole bunch of pushback recently decks, who did this great piece on like Twilight factories versus the dark software factory you put in what it should do and how you know it works. And it does the whole thing. It writes the code, it verifies it, pushes it to the point.

That's a dark factory. Why light factory turns out you need to leave the lights on a little bit for a few humans to hang out because it doesn't do all the things with a level of judgment we need. There's another talk recently, exoskeletons, not factories. Maybe what we want to do is make that Ironman suit that makes us more productive. All of these are valid. It doesn't really matter. But the more you try it, the more you'll find where it breaks, where you have to bring in the human in the loop judgment and how you can accelerate the rest. So that's so good. Let me let me let me.

I I I want to get to another question just just because I think it's super important. You obviously lead lead a community called gather.dev. I'm the head of community at Torque, which is, you know, obviously community first developer platform. There's one question here says community as developer career mode in a world where coding syntax and boilerplates can be generated by LLM in seconds. What actually gives an engineer long term career durability? So can you talk about why community is more important than ever? Yeah. So what gives you long term career durability is a combination of things. Some are and are not related to community. I think the first thing is agency.

The robots can do most anything, but they won't unless you tell them to. So if you can just have agency, if you could just step up and say, this is something I want to see in the world, that is one of the top things people are hiring for. It ties in very tightly with curiosity. Can I do that? Would I do that? What happens if I do that? That experimentation shows that you've got the passion to figure stuff out and to push the envelope, which is incredibly valuable. The second part is you need to be able to learn more quickly than anybody else or than other people.

Right? I want to hire the person who can learn the fastest, consistent with them, also having the time and engagement and energy to deliver high value, engage with the customers, all the rest. So one of the things that community gives you is a network of people you can learn from that you have trusted. I always used to like back in the day, I had a buddy where I'd fix his sequel queries. He'd fix my jQuery. Like I wasn't so good at front end. I was a little better at back end. I was like, yeah, right. This was a while ago. This was like 20 years.

I love that. But that's it. So community gives you a place to learn, a place to see your unknown unknowns and a set of people you trust and like to engage with, to continue to improve your understanding as things change. We can't figure this all out alone. There are too many experiments to run. Having a team of people to run them with will accelerate your career success. I love that. Let me let's let's kind of get into let's kind of get a few more topics. To be honest with you, I think we may do a little bit more rapid fire because these questions are really good.

I know these are fully loaded and maybe we should bring you on for part two, man. So I want to talk about what engineering leaders actually screen for in 2026, right? So again, my your ICP is hiring leaders. My ICP is a talent. I feel like there's something there. But talk, talk about if people are watching this list is now in the future. How are engineering leaders screening right now based off your conversation? Again, this is just Peter. I always, this is just one sliver of the internet is what I say. So, funnily enough for the majority, I speak to a lot of CTOs, hundreds a year, maybe low thousands.

The vast majority of them are still screening the way they did a year or two ago. And that's failing. The ones who get it are screening for agency, curiosity, proof of work, flexibility and a passion for either factories and agents and or solving business problems, solving customer problems. It's that passion, that engagement, because nobody, I don't have the time to manage somebody. Look, I want to do one on ones. I want to care about their grandma. And if they're cat sick, like that's all important and valid stuff. I don't want to have to tell them what to do on Wednesday. I need them to go figure it out.

And so I'm primarily screening for that. And the easiest way you prove that blog posts, videos, open source projects, it doesn't matter, get proof of work out there, proof of curiosity, proof of agency. That's what differentiates. I love it. I love it. Um, let's see here that the. Uh, I'm just going to go through these global global town versus autonomous agents. You look at engineering or adopting AI to boost velocity. Well, across global town networks, we see companies combining AI tools with distributed global engineers and executives have a C run, blah, blah, blah. Uh, our CTO is viewing AI to replace the distributed global engineering or as an accelerator.

We kind of talked about this a little bit, but is there anything else you like to add there? All of this stuff is yes. And there's always been people who used global talent as a way to reduce costs. And it can much more importantly, just like being a little more flexible about whether or not somebody went to Stanford or not. Maybe they went to a boot camp, but they were an awesome programmer and they live in, you know, Mississippi, like that's great talent too. Uh, geography is another way to get more great talent into your company. Uh, then there is a separate dimension as to, okay, once you've got great talent, how much, what's the ratio between talent and agents and what's your ambition?

Like if you've got unlimited ambition, Shopify is hiring interns and they're using agents. Bloomberg continues to hire interns. It's a great time to hire junior developers. And at the same time, all the companies would just be like, I would just keep 30% of our seniors and cut the rest that they haven't done it yet, but they will. Yeah. Okay. Um, let's see here. I think I think we covered the rest. I think we covered the rest. All right. So, uh, anything we haven't covered, uh, that's kind of been sitting in the back of your brain before I ask you some rapid fire questions. I think definitely I do want to just double down on that future of the junior engineer because it goes back and forth and back and forth and it keeps coming up.

There is no question. It's a harder time to be graduating from, you know, some kind of engineering degree or a boot camp than it's been for a long while. You do need proof of work. You do need proof of agency. We are going to find one of the best pairings, pair programming has always been useful. If you compare a senior who knows your environment and, you know, where the bodies are buried with a junior who's super AI peeled, they will both learn and prosper for that as they pair while engaging with a set of agents. So I think companies are being very shortsighted if they're not investing in junior engineers, not because we want to worry about where the seniors come from in five years, but because it's a cost effective way to bring new ideas in an AI pool, the company and quite frankly now to get incredibly smart people at a very fair price.

Yeah, I agree. Another thing to this actually just brought up a really good kind of thing that's been sitting in the back of my brain. I've, you know, the, the, you know, we obviously do within our community. We do resume reviews, tort profile reviews, you know, LinkedIn reviews, report, you know, we do all that to help people, right? You know, the biggest question I get time and time again is, you know, how do I separate myself right now? Like, listen, I am fully AI peeled. Like, like we've talked about there's no reason that you can't spin up some sort of product that can be used by a handful of people.

Like there's just no reason. There's. And so, you know, if you're on the job search and, and, you know, you've been on job search for eight months and I ask you as a recruiter, where have you been doing a for a month and for eight months and all you say is self-study. I'm not going to fly. So thoughts on that. Absolutely. And you can get so creative with this, right? For many companies, you go through LinkedIn, you can see who the hiring manager might be. I just, I mean, let's be creepy about this. Go do some research about what they're passionate about and then send an unsolicited LinkedIn pointing them to some kind of web app you built just for them, like for something they care about.

It's like, dude, put this together in the evening. I'd love to speak to you about how I could use AI to accelerate, you know, development within your team. You basically, the process is always if you want to get a job, there's two things you need to somehow do. One is you fill in the resume and gold fill out the board at Greenhouse or wherever you're doing it. And then the other is figure out how to back channel in cold or warm, get into their DMs. Can you find them on Twitter? Can you post authentically, positively and usefully deliver value to the person who you'd like to hire you?

And then that's the way that you potentially break through the recruiter screen. The one other high agency way, and this is really hard to do, is you become the person who is the intersection of two things, find two things you care about, one of which exists, one of which is new. I was doing domain specific languages in Groovy, a dynamic type language on the JVM or Ruby back in the day. And most people didn't talk about it. So I got just presented all the conferences. Again, I had the luxury of being able to travel. I had contract work. I could make it work. I didn't have kids back then.

So it's these are ways that if you can become the voice, X is designed to find the next big thing, become the next big thing on a niche topic, the future of using storyboard to build domain specific languages that can accelerate the development and regeneration of front ends for corporate websites. Because somebody's going to need some really AI-filled front end person and then you filter just for that. The more you can, if you're just like, I'm going to apply for any programming job, you're going to send 20,000 resumes and get zipped. If you're like, now, I'm only going to do series B, series C, B2C, primarily cosmetics front end design or, you know, series D plus with really gnarly data problems because ask me about Kafka or in stream processing.

That's the place you need to try to build that confidence. You don't have to be the biggest voice. You just have to be one of them. Yep. I love it. So good. All right. So I got some underrated. I asked my AI. I asked my robot. I said, Hey, I need you to come up with me. Can we come up? Yeah. I said, overrated, underrated. Give me some questions. This is, this is a, this is just a pick one here. Okay. Ready? AI code auto complete cursor or co-pilot versus autonomous background agents. Which one? Autonomous. Okay. Yeah. It's not 2024 anymore. Come on. Isn't it crazy?

That's the way I feel. I'm like, I'm not using cloud as much anymore. It's crazy. No. All right. Overrated, underrated prompt engineering courses for software engineers. Bizarrely the exact words you use matter, but much more important is the multi step processes, verifications and loops. So it's necessary, but not nearly sufficient. Okay. Overrated, underrated. AI will replace junior devs in two years. Uh, no, some junior devs will be replaced by AI. Uh, they will continue to be a role for, uh, junior devs and AI. AI generated PR summaries and code reviews. Depends upon how well you build them. The same as the guitar is a good or bad depends on who's playing it.

But they're great. If you do it right. Okay. Heavy agent frameworks, Langchain, crew AI versus building your own simple loops. You'll need the rigor over time. You learn more by building your own system. You should be capable of both. Okay. Tracking lines of code or PR velocity to measure team AI adoption. As long as you don't use it for performance review, it's one useful data point. If you're looking, look at cycle time, PR's per engineer, and then, um, like meantime, between failure, something about your change failure rate, that way you've got a compensating control. So you don't just go faster and break more stuff. Love it.

Massive context windows, one million plus tokens versus targeted rag. It's all about targeted context, not necessarily just rag. You've got to put the right information in there. They got a million tokens. They miss most of them. There's a whole bunch of retrieval stuff that missing in the middle. Once you hit 40% of context, it's time to start new session. Okay. Very good. Fine tuning smaller models versus prompt chaining frontier models. Interesting. Depends whether you are going for you prototype and iterate with frontier models with prompts and multi step processes. You then eventually, if you have sufficient volume, consider fine tuning a smaller, cheaper model to reduce latency and cost, especially at enterprise scale.

Love it. AI generated test suites and synthetic test data. Synthetic test data is one of the pieces. You're going to have AI generated test suites. You just need to make them better. Verification of correctness is a huge problem. You need to do all the things, defense and depth, just like security. Love it. Last question. The concept of the one person billion dollar company. It's it's coming. The reason to have companies with large numbers of people to reduce transaction costs and trust costs, all of those are externalities and problems that can be solved agentically and with technology. It will take time. But eventually, we're going to see, I think, a lot more smaller companies having a bigger impact with more informal interaction mechanisms through agentic contracts.

I love it. I love it. Awesome. Awesome. All right. Last question. I always ask people here. The shows called guidance counselor 2.0 because we had guidance also, again, shout out, pally.so for those questions. The shows called guidance counselor 2.0 because we had guidance counselors in middle school, high school and college, but now as adults, we don't necessarily have guidance counselors return to to ask for career advice. What if you could go back in time until young Peter, one piece of advice, what would it be? I mean, it's very specific. I went to the college that was best for the thing I wanted to learn, rather than going to the best college, could have gone to Cambridge.

I chose not to, which in the UK is a pretty good college. I'm sorry, I should have done that. Turns out, you know what? I would generalize that lesson to the people, Madam Moore, the details of the topic. Yes. Same reason you need a community now. It's the community of alumni, community of people you work with, community of people you join. I love it. I love it. Well, Peter, how can people find you, man? Let's go and wrap it up here. How can people find you? So they can check out the website, gather.dev. And then I am just Peter Bell on X. That's the easiest.

I love LinkedIn connections. Unfortunately, there's a limit of 30,000 that I'm at it. So I have to remove somebody every time once. So I got like a backlog of 700 LinkedIn connects. So connect with me on X. And it's gathered at gather.dev is you want to drop me a line. I love it. I love it. Well, y'all, thank you again for everything. Thank you again for tuning in this morning. Hey, you did it. You listened to an entire episode and I am very grateful. Because you know why? There's a lot of podcasts out there. There's a lot of episodes and I myself rarely finished an entire episode.

So I appreciate you so, so much for listening. Quick reminder, please rate and review this podcast in the app store. I want this to be discoverable to others. And the only way that happens is if you take the time right now and rate and review it, please, like right now. Unless you're driving, then don't do it. So thank you again. If you need to reach out to me at T. S and on all social media handles. And I hope you have a fantastic day.