Beyond Vibe Coding with Addy Osmani

What does by quoting mean to you? By coding is not the same as AI-assisted engineering, and I feel like that distinction is kind of critical, because we don't want to devalue the discipline of engineering. How do you personally use these tools? I have more recently been focusing on the idea of spectrum development, having a very clear plan of what it is that I want to build. If you are open to tests, it can be a great way of de-risking your use of LLMs in coding.

One thing that I'm kind of noticing on myself already is losing a little bit of criticality. It's going to continue to be very important for us to be able to think through how things work, be able to problem solve without necessarily relying on the AI testing, and retesting your critical thinking skills are going to be important. You're working inside a larger team of the Chrome team, other teams as well. What are things that you're observing at a company like Google, with AI, what we've realized is how do professional software engineers at the likes of Google go beyond vibe coding speed up their days they work with AI. Addy Osmani has worked on the Chrome team for 13 years, and if you ever opened up the developer tab on Chrome, you've definitely used the stuff he built.

He is also a prolific author. His latest book is titled Beyond Vibe Coding, and it's aimed at professional software engineers. Today, we go into vibe coding versus AI-assisted engineering, and why vibe coding isn't useful for much more than just prototyping something quick and dirty. The importance of understanding what the model does, why Addy always reached through the thinking log of the model he uses, so make sure he fully understands what it did, and why, before approving changes. New development workflows with AI, how things like speculative and development, asynchronous coding background agents, and parallel coding with several agents are new and unexplored areas of software engineering, and many more.

If you're a software engineer who wants to work better with AI coding tools on the day to day, and wants to build a reliable software with AI tools, then this episode is for you. This podcast episode is presented by Statsik, the unified platform for flags, analytics, experiments, and more. Check out the show and also learn more about them, and our artists as sponsor. So Addy, welcome to the podcast. Thank you.

It's a pleasure to be here. You're an author of a lot of different books leading effective engineering teams that came out about a year ago, and now your latest book is called Beyond Vibe Coding. And on your sub stack, you also write a lot about your learnings about vibe coding slash AI assistant development. In this book and also on your blog, you talk about vibe coding versus AI assistant software engineering. In your mind, what does vibe coding mean to you specifically?

And how is it different or similar to AI assistant software engineering? Yeah. So I've been a ton of tens of folks that I personally think that vibe coding is not the same as AI assistant engineering. And I feel like that distinction is kind of critical, because we don't want to devalue the discipline of engineering and give folks that are new to the industry and incomplete picture of what it takes to build robust production-ready software. But I kind of have, I guess, two definitions for them.

I think the vibe coding is really about fully giving into the creative flow within AI, so very much focused on high-level prompting. And in many ways, forgetting that the code exists. So this goes back to Andrze Karpathis, like original definition. But it's about accepting AI suggestions without necessarily having a deep review and focusing on rapid iterative experimentation. Personally, I find it really great for prototypes, MVPs, and learning.

And I think it's what I've seen in production teams is that this has been very useful for them when it comes to trying out ideas rapidly. And building out an intuition for what the shape of an idea might look like, what a component might look like, what an MVP might look like. And that tends to prioritize speed and exploration over things like correctness and maintainability. Things that we perhaps care a little bit about when it comes to building things for large production audiences. And I would say there's a little bit of a spectrum between vibe coding and doing what falls a little bit closer into traditional software engineering, doing more planning, more specification driven development, including sufficient context.

And really what is AI-assisted engineering across the full software development lifecycle. So to me, AI-assisted engineering is where AI is this powerful collaborator, but it's not a replacement for engineering principles. And so in that model, it is a force multiplier, but it can help you across that whole cycle, whether it is with boilerplate, debugging, deployment. But the big difference is that the human engineer remains firmly in control. You are responsible for thinking about the architecture, for reviewing the code and for understanding every line, if not most of what AI is generating for you.

And you're on the hook really for making sure that the final product is secure, scalable, and maintainable. Maybe AI is helping you with velocity, which is great, but that doesn't mean that you can sort of shrug off your responsibility for quality at the end of the day. And I tend to think that a lot of people have said that they found AI encoding to be a force multiplier. But I found that the greater expertise that you have in software engineering, the better the results you can get when you're using LLMs. And I think sometimes if you're new to the industry or you're a junior, maybe there are some, what we would consider to be traditional best practices that maybe you haven't had to experience them yet or think about them.

Like, if you care about production quality programming, you should probably only be committing code to your repo that you can fully explain to somebody else. Because just expecting that the AI is going to help you untangle whatever mess happens later on is probably not going to work out on long term. And how do you personally use these tools? And may that be vipoding, especially AI-assisted engineering? I have more recently been focusing on the idea of spec-driven development, having a very clear plan of what it is that I want to build.

I think that there are definitely places where I still vipecode. If it's for a personal tool, something through a way, or in the old days, if an engineer or a PM had an idea, maybe it would be put together a quick mock, maybe we would put together a wireframe or a sketch or something like that, perhaps work with UX to come up with something a little bit more polished. These days, the fact that you can vipecode a prototype and actually show somebody in a pull request or in a chat, like, hey, here's an even more clear version of the vision that I had in mind for this. I think there's something very powerful to that. And I'm loving vipecoding for that, just being able to give me a slightly higher quality way to deliver sharing an idea.

Vipecoding is having its moment. Describe an app and a prompt and boom, you've got something running. It feels like magic. But here's what vipecoding doesn't account for. Institutional knowledge, every prompt comes from what you tell it, not what your team already knows.

Real product development is the opposite. It's accumulated context. For example, everybody has a kind of history. Every feature connects to customer requests. Every PR fits into your team's roadmap.

Short prompt will not share all this additional context for the agent to use. Linear agents work inside this additional team context. They live inside your development system where actual work happens. They see the issue blocking your sprint, the really if PR is the project goals, the discussions that your team already had about this exact problem. And because linear is your team's shared workspace, agents don't only see your context, they see your entire org's context.

The blog report from support, the design spec from your PM, the architecture knows from the tech lead. So when you ask an agent to draft a PR or build a feature, it's not improvising just from a prompt. It's using the same context that your team already uses. This is what AI-powered development looks like when you're building something real, not just building fast but building intentional. See how agents work in linear at linear.app.agents.

And if you want to try them yourself, get linear business for free by visiting linear.app.app slash pragmatic. That doesn't mean that we take the vibe coded prototype with that code and just stick it in production. Once you have clarity around the vision that you have for a component, for a feature for a view, you should probably be writing out like, okay, well, what are the actual expectations around this? Which we actually consider to be the requirements. And that will give you a much higher quality outcome typically from the LLM.

Because otherwise, if you're giving into the vibes, you're also sort of giving into, okay, well, you figure out the architecture. You figure out like what they should do. And well, that's fine for ideation. It's probably not sufficient for production sort of product engineering. So for me, spec driven development is a big thing.

I think that tests are great. And if you are open to tests, it can be a great way of de-risking your use of LLMs and coding. Because sometimes, even if you're using state of the art models, you can end up in a situation where maybe the code looks more convoluted than you would expect. Or maybe your first couple of prompts have been generating really good code. And then for whatever reason, things go off the rails.

But if you can prove that things are working with tests and that if something did go off the rails, it's clearer to you what that was. I think that can help you keep your project green the whole time. And that's been something that's helped me quite a lot as well. So spec driven development, testing. I try to also make sure that I'm leveraging.

I guess this is a shout out. My team just released Chrome DevTools MCP. So I use a ton of like. Oh, I'm seeing. Yeah, so we just released Chrome DevTools MCP.

I care a lot about quality. And I think that in the last couple of years, we've seen a number of cases where, if you notice that something is broken, I will see a lot of engineers that will just say, OK, well, hey, LM, it looks like this button is off. It looks like this thing is not quite exactly what it should be. Go fix it. And that's again going a little bit closer to just rolling with the vibes.

But if you can keep things like a browser in the loop or something that can actually see the page, that is what Chrome DevTools MCP and related solutions do. They give your LLM eyes in many ways. So we can see what the browser sees. They can see what's rendering or what isn't. It can detect if there are warnings, errors in the console, maybe even get a deeper sense of what is broken.

And that can just improve the feedback loop that you have. And so I've been just excited about MCP for being able to help us evolve our workflows a lot more, thanks to this idea of being able to call in to other tools. So that has been great. Otherwise, for me, generally speaking, I've found that you really do need to put in the efforts to get proficient with these tools. And I still find myself doing that.

If there are new models, new tools, new platforms coming out, I'm generally finding that I experiment a lot every week. I try to encourage my teams to share with each other with me, like how are things going? Are there any insights that are worth us bubbling up or things that we should try out as a team? And if your team see that you are very open to learning together, I think that that can just create this nice culture of psychological safety that can set your team up for success as we're all going through this big period of change. Yeah, and I feel what you said about it takes time to learn this thing.

I've had so many smaller and larger moments just by using it. And there's not many people who are skeptical, especially including engineers who are skeptical about LLMs, whether it may be the theory, may that be energy footprint, other things. But I found that a lot of people who are skeptical have either not tried it or they haven't given it some time because it does take some time, some playing around. And you know, like, we're engineers, it doesn't work everywhere. It makes mistakes, it screws up.

But it does, like, I think if you've not felt like, I can only tell for myself, but I found a bunch of ways where it helps even my smaller projects. And as you said, like, it's about what works for you and what works for others. Speaking of which, you're working in a larger, inside a larger team, larger production team, you've got the Chrome team, other teams as well. What are things that you're observing on how others are using it? And interesting ways, maybe unexpected ways, or even ways that maybe didn't work out for those specific people or engineers.

I would say that at a company like Google, we have this very long-term, very battle-tested way of thinking about software engineering at scale. And with AI, what we've realized is, a lot of that doesn't go away. You still want secure about quality. You still want to care about doing due diligence. The things that I think we've found interesting kind of mimic what people have seen in startups and other kinds of companies.

So the importance of mastering understanding, like what is prompt engineering, right? Like making sure that you are constructing the right set of incantations to get the best outcome from an LLM. And then context engineering more recently, like how can you make sure that you are optimizing the context window to increase the chances that those outcomes are higher quality? We spend a lot of time thinking about that, making sure that the right sets of descriptions, details, files, examples, any additional content that is specific to a project that the LLM may not necessarily have in its training data. That has been very interesting and important for us as we've been working through.

I have been somebody that has been trying to explore using AI in every facet of my life in many ways over the last couple of years. And that's been very eye-opening in terms of places where there is meaningful productivity gained to be had, as well as places where either model quality or system prompts and tool quality are not quite there just yet. And so I've been also sort of nudging my teams in that direction as well. If you are thinking about this idea of us eventually all being AI native engineers, then one prompt for you is before you try to solve a problem yourself, if you gave this to a model, you gave this to AI, what would it do?

Would it actually help you accomplish your goal much faster? Or would it likely slow you down? And if it would slow you down, why is that? But even that prompt is something that I think helps us to learn a lot about what is possible and what isn't. Many of us have been on this journey where, if you're a classics software engineer, if you're a web developer, whatever, you probably don't have this very deep understanding of AI.

For myself and for some of my leads, I set us this task of like, okay, well, let's start to become slightly deeper experts in these areas so that we can guide our team towards, well, where would it be useful for you to build up this expertise too? So again, in the last year, I've been spending a lot more time on thinking about things like EVALS and benchmarks and how much are we gonna be caring about, RAG versus fine tuning and so on. And this ends up contributing as well, because at the same time that we are talking about AI assisted engineering, many of the products that we work on also have this consideration for well, is AI gonna help you deliver a better customer experience in some way? And so I'm trying to look at like, how can the work that we're doing for our coding workflow also benefit the product work that we're gonna have to do?

So that's been a good learning for us. I think honestly, just the importance of always maintaining human oversight has been one of the bigger learnings. We have, of course, like I knew a lot of people have run into this. We've of course seen cases where people, very often like external contributors, will sometimes be very passionate about like, hey, I wanna contribute to your project, but then they'll use an LN. Where you're going with it?

Use an LNN, something that I think Michelle Hashimoto was blowing up on social media because he just had enough that people were submitting things. And as you said, out of good intention, but putting way more load on maintainers. I love reading studies about how different segments of the developer population are finding AI and what that means for their teams. And what are the studies that I recently read, highlighted that if you are increasing the velocity of code and improvements that can land in the team and you are doing human oversight, the human review is gonna become the bottleneck. And that is what teams are starting to realize, like, oh, wait, we're starting to see a lot more PRs come through, but who's gonna review those, right?

So I'm glad that teams are, at least some teams seem to be caring about that quality dimension sufficiently to actually hand review themselves, but that does also mean that our workflows may have to evolve. I have seen a lot of talk about like, yeah, why aren't we using LMs to also do the code review? And that's a little bit of a slippery slope because if the AI is writing the code and you haven't studied it carefully, but the AI is also reviewing the code, are you actually sure about what's landing? So I think that the best practices around code review are still something that are evolving. And I'm excited to see where that goes.

For me personally, there are a lot of tools that I use, some of my favorite ones include using Cline in VS Code, a lot of people. You're a big fan of people following your writing on Cline. Yeah, I love Cline. I think that people can do a lot with cursor and co-pilot in VS Code these days as well. But one of the things that I enjoy doing is, most of these tools will show you the thinking that is happening behind the scenes as a solution is being built out.

And even if that happens quickly, like I will try to go back, scroll through, expand and read through what was your thinking process through actually being able to build this out. What decisions did you make? What did you generate? And I will review that code before it ends up in a pull request. There's going to be some likelihood later on that I may have to maintain that code, that I may have to make some tweaks to that code that the L on is not going to be able to help me with.

Just last night, I was working through a problem. And the code looked, you know, on paper, code looked right. It wasn't working the way it was supposed to. I asked the L on a few times to like, hey, can you go user tools, go figure out what the problem is? It continued to make changes, didn't actually fix the underlying problem.

And so tonight, I'll have to roll up my sleeves and go manually debug it. If I didn't understand how that code worked or I hadn't been reading it, I would be feeling like I'd just been dropped into a jungle and haven't been navigated myself. Well, plus, I mean, I'm reflecting on this, right? Like I feel this is the difference between a software engineer who's worth their money, just in terms of, you know, being employed. And one, that's not, not if all you can do is just prompt and prompt. I mean, anyone can do that.

I'm exaggerating, but a new grad can do that or someone's still in college. But not everyone will have taken the effort to understand to know how it works and be able to roll their sleeves. And when these models fail, which they fail, they can fix it. And also they can also explain it to people, they can explain it to the meeting coherently without rambling. So I feel, you know, like, this is what a professional is, right?

Like in any industry, it's like they know how the things work. The car mechanics, they mean with anything else. So maybe it's just a reminder that if you're not doing this, if you're kind of letting go and like, oh, it solves my problems, you're kind of at the risk of, well, if you don't know how it works, I mean, why do we need you? Anyone can prompt another LOM. Exactly, exactly.

And, you know, as we start to talk more about, you know, in the last year, you know, people working their terminal, their CLI has become more of a thing once again with cloud code and Gemini CLI and open code and so on. And then people started talking about, okay, well, how do we orchestrate multiple parallel agents being able to complete work for us? As we start thinking about this idea of each engineer almost having a virtual team of their own being able to go off and work through your backlog and get all of these different tasks implemented concurrently. I think that you start to quickly realize, well, that all sounds great in the abstract and in theory, but then you're gonna compound all of these other problems where a lack of manual review is going to probably lead you to some level of tech debt, maybe not immediately, but at some point, it very likely will.

And my experience of this kind of led me at one point to write about what I call the 70% kind of problem. Let's talk about that because yeah, you wrote about the extensively, we actually published a shared article, I guess, article based on your article, which we'll also link in the notes below. What is this 70% problem? And how did you come across it? And how do you think it might have changed since you published it, which was about six months ago?

Yeah, obviously model quality to link quality continues to get better. So the 70% problem is really about this idea that LMs can produce very roughly like 70% of a working application very quickly, but they tend to struggle with that last 30%. And that last 30%, you can consider it like the last mile, but it includes lots of different kinds of patterns that your audience probably run across, or maybe they will run across, things like the two steps back pattern. So you have used a couple of prompts to build up something, you've given LM one more prompt, and it happens to go in a completely different direction. Maybe it is fully rewritten, your UI, or the functionality behind your component, things like that.

They're very often hidden maintainability costs, places where you're not being specific, you are delegating the responsibility back to the LLM, and you may end up getting diminishing returns. As we've seen time and time again on Hacker News, security vulnerabilities. This is exactly where we see people accidentally leaking API keys where there are access issues, where there are all kinds of problems, because people did not think as holistically about the problem that they were solving, and just sort of gave into the vibes. So a Vicodid proof of concept, it's fine for MVP for that prototyping phase, but it likely needs to be rewritten with production quality in mind, if you are going to be landing this in a code base, where you're working with other people, working with a team, and dealing with a real user base of people.

I think that the security and quality aspects really speak to the need for keeping the human in the loop there. Yeah, and we keep hearing stories about product managers and less technical founders who get really into vibes coding excited, they spin up a prototype or a better version, and then they want to build something that they can put out into production, and there's a bunch of stories I'll link one in the show notes on how they just get stuck or it takes them a very long time, what they thought would take a day or two, ends up being 10, 20, 30 days, and in a 70% problem, you went into something relevant for this, which is you said that more experienced engineers can finish the last 30% a lot easier, and then less experienced engineers get actually a lot more stuck or get this false confidence.

Yes, yes, absolutely. I think that with that last 30%, what you will often see junior engineers or interns or new grads doing is they will not really know what to do next beyond just constantly re-praunting the LLM to fix the issues, and if it's not able to do it, they're not necessarily gonna have all of those skills just yet for debugging the problem or understanding where the issues are. And so what this speaks to is the importance of having a really good critical thinking, problem solving mindset. We always talk about the importance of this for people who are going into computer science. I think that remains now, but that diligence requires to actually read the code, understand the system, understand how all of those pieces connect together.

I don't think that that necessarily goes away. As you said earlier, with the tools that we have today and the models we have today, almost anybody can give a high level prompt to a tool and have something, you know, that seems like it works, come out the back of it. But I wouldn't necessarily trust that in production. Too many stories at this point of things just going a little bit off the rails. It reminds me when this was before AI tools, I had an engineer who joined Uber, who came from a smaller startup, and I, a new joiners after a week, I was this person manager, I asked like, how are things going?

And he was really distressed, I'm like, what's difficult? He's like, I'm trying to read all the code, and it's just too much. And I was like, why are you trying to read all the code of Uber, which was like, you know, our back-and-sys, and it was like, you know, like, just so it was more well. It was like more than me, like, like, well, I do that because whenever I join a new company, I read all the code to understand everything. And I was like, I give what you're coming from.

And I think that's great. I was just like, let me, let me explain to you how the code base works is you shouldn't read all the code, but you should understand the structure, you know, where you can find things, because this is, you know, this is a very useful base where we had like 200 or 300 engineers working on it. And I thought this person had a really great intention, which is I'm coming to a new place. I want to understand it. I'm going to spend the first few weeks understanding and those people excel.

And I'm kind of thinking, we keep getting back to this thing. But what I'm hearing from you is, is if you can actually, to some extent, outsource this to an LLM, which is a bit of a hit or miss, it might or might not succeed. But if you do, you're now hopelessly dependent on it, and at some point, when the context window fills up or when the model is not having a good day, because these things are non-deterministic, you're kind of stuck. You know, your best thing is to try a new model or empty the context window or I don't know, do something else. But it doesn't really feel like you're in control.

Yes, yes, you're absolutely right. And I think that whether your gateway into this new world was vibe coding, where you are a senior engineer and you've been evolving your workflow for AI, there are a few things I think that everybody should keep in mind that can help you get the best outcomes. Addy just mentioned getting the best outcomes. Here's what I'm seeing with AI in vibe coding. These tools can give you incredible velocity.

You can ship features a lot faster than before. But velocity with that precision means you're just shipping more things, not necessarily the right things. How do you know what you build actually works? Did that new checkout flow improve conversion or heard it? Is it feature you should help you retention or cause a drop off without precision in how you roll out a measure, you're making decisions blind.

That's where Statsik comes in. Statsik is our presenting partner for the season and they built a precision toolkit that matches your AI accelerated velocity. Here's what precision looks like. You ship a feature to 10% of users in a controlled environment and see if it's moving your metrics up or down. If conversion drops, you catch it early and roll back before it affects everyone.

You're making precise, data-driven decisions at the same pace you're shipping code. Take Graphite as an example. They build granular control and rapid iteration into their development workflow using Statsik. When they roll out a new feature, built an analytics show exactly how it affects your key metrics. They can see if engaging this up, if the feature is causing errors or if user is dropping off.

They do do so feature gates and metrics working together and they're running over 300 of these controls while outside any given time. During production incidents, this approach cut their resolution times by over 50% because they could quickly identify which feature flag was causing issues and fix it instantly. Most teams stick together separate systems, wait on queries and try to correlate user segments that don't match. By the time they know if something worked, they've already moved on to the next feature. With Statsik, you have everything in one place.

Feature flags, experimentation and analytics, with the same user data. Statsik has a general street here to get started and pro pricing for teams starts at $150 per month. To learn more and get a 30-day enterprise trial, go to statsik.com slash pragmatic. With this, let's get back to the conversation about AI and development workflows. Understand that models tend to have finite context windows.

They've been getting larger over time. You probably need to at some level adopt sort of a project manager mindset. Break tasks into small verifiable chunks. I've seen a lot of people who will throw the kitchen sink at the LML and say, hey, yes, build all these requirements at once. And that doesn't necessarily work the best.

So you're small verifiable chunks, clear expectations and being prepared to iterate with the AI. Don't just feel like one-shotting is gonna give you the best outcome because it probably isn't. And that decomposition is really pretty similar to planning a sprint or writing pseudocode, the type of thing that we would do in the older days. And it reduces the risk of context loss and compounding errors. I think that a lot of best practices and software engineering remain timeless even with AI coding.

So caring about modular, testable code, enforcing those code reviews, AI is gonna introduce a few new habits, like thinking about the input and output constraints and seating enough context. But that doesn't mean that you can't get good output from it. It does mean that you're gonna have to apply a level of diligence with that human in the loop to make sure that you are setting yourself up for success. The more that you give up your responsibility so that the other land, the higher the risk is that something's gonna go off the rails. Yeah, and one thing that I heard from an engineer, actually, I'm in one-ature who's a longtime center engineer, the creator of Flask and the Bunch of Open Source frameworks.

And he was telling me that he's observed something interesting with engineers who are firmly in control and very confident in their work. He said that he sees they're just seeing a lot more success with AI. And the people who feel that they don't have that much control over their work, over their tasks, their design tickets, and they also have a bit of the world's control. And you might say they're freaking out a lot more about AI and what if we'll do. And this was so interesting of, it just came to me as we're talking about it, because what I'm sensing from a conversation is, again, a lot of your advice boils down to be in charge, understand everything, be confident, that if this thing is taken away, you can make progress, like no problem, it'll just be a bit slower.

And as long as you have this mindset, it feels like everything is easier. And keeping this, like you know, as you said, like you still read through the thinking, you prefer the models where it explains to you, and then you can go back if you wish, and you often do it, you read through to make sure, so it's kind of fun. You keep learning professionally, you keep getting better every day. Absolutely, absolutely right. I'm reminded of, you know, AI is just another tool in your tool belt.

You know, we've gone through so many different moments over time of engineering and developer experience getting a little bit better with every generation. You know, when I was coming up, I remember getting excited when templates became a thing. You could generate code with templates, you mean? Oh, no, just downloading templates for ideas for UI for, you know, starting off. Oh, on the web, right?

Yeah, on the web. And this was just the simplest of things, it was like a zip file, right? Somebody also created it. Okay, so I've improved my starting point. And then we got stronger command line interfaces, scaffolding, that type of thing.

And then you didn't have to worry about the exact starting point. And I feel like this is just taking us a few steps forward. Like it's making it easier for us to bootstrap a solution that kind of works, but still requires you to really understand what you are putting into your code base shipping out to users and not, you know, getting rid of your responsibilities at the end of the day. I find that, at least for me, in the last year or two, going back to first principles, as always, like, has helped a lot. I've been working a little bit more closely with the Google DeepMind team on Gemini and its coding capabilities.

And that has been a really good reminder for me of just understanding how a lot of these things work behind the scenes. If you remember that, you know, what is training data? Well, it's very likely gonna be looking at permissively licensed code that's on GitHub or out on the open web. The patterns in that code are probably gonna be reflective of, in many cases, like lowest common denominator, things that maybe just work, are they gonna be the most secure, the most high performance, the most accessible, possibly not. And so if you remember that the training data itself still requires a lot of work to actually get things to a place where, you know, it maybe is a production quality.

You sort of sort of set your expectations a little bit lower when working with LLMs. You realize, like, yeah, it will probably do a better job than we just copy and paste in snippets of code from somewhere else, obviously. But I still very likely need to do a level of manual work or a level of diligence on top of this to get the best outputs. Yeah, it reminds me a little bit of 10 years ago or so, 10, 5, 10 years ago. We use a lot of Stack Overflow because when you search for something it was there.

And so for example, for email validation, like you would search like, how do I validate email address? There was a Stack Overflow question and it had like 20 answers, one of them was top rated, it was a regular expression. And what I did, and I think what a lot of people did, is I just copied and pasted because I cannot be bothered to define exactly the regular expression of emails, which is a lot more complicated than just things. So I just kind of, if it was for nothing important, it kind of worked. But there were then complaints and flags raised that a few of them, not this is an example, but they were insecure or they didn't account for edge cases.

And a lot of devs, again, when you were in the mindless mode or it didn't matter or you know better, you did that. And I guess we're gonna have the same thing at larger scale. Like, why did the bug happen? And you're gonna look back at the history. Yeah, someone just looks good to me for the big four requests that was Indian generated by an AI.

I think that this is very much a smaller point, a side discussion point, but I've seen a number of cases where people will now say, hey, do I still need to use third party libraries if I can just prompt a very slightly smaller version of a solution myself, which reminded me exactly of the stack overflow copy-pacing, the top-response thing. And if you were a senior engineer, which you realized is, well, hold on, that also means that you are now taking on the responsibility of making sure that this is future-proof for security issues, for all sorts of platform issues that could happen. If you're relying on a library, it's perhaps easier for there to be a central point of leverage where those fixes can be made and then deployed out to people. If you own all of these different patterns in your code base yourself, that also means that you're taking on the responsibility.

Can be totally okay, but it's something that I sometimes see people just not being mindful of requiring a little bit of extra work. And I feel this is where we just need to remind ourselves that that is just engineering, right? It's trade-off. Do you take on the responsibility and the risk of maintaining this thing and the fact that it might have missing edge cases or whatever, or might not work correctly or securely, or do you take on a dependency, which has its own problems, right? There's now, yes, dependencies is curious, et cetera.

But speaking of the need for software engineering, what are some new workflows, some things that we have not been able to do before as software engineers that we can now experiment with? You might be trying out with these AI tools that it's just like, is brand new. Yeah, I think for me, the thing that I'm most excited about seeing evolve is this idea of asynchronous background coding agents, the number of teams playing around with these ideas at the moment, Jules, Codex, we also see GitHub experimenting with some of these ideas as well. I think that the idea of you being able to delegate parts of your backlog and have a system be able to implement that in, you know, asynchronously is a very interesting idea.

If we can do that without merge conflicts in a way that is easily human verifiable, so again, going back to the human in the loop, I think that's very interesting. I have been finding that that is very, very, very much already in a place where if you are having, you know, one agent work on writing or updating your tests, or you have a number of agents working together on trying to migrate your code base from one version of a library to another, or, you know, a version of a dependency to another, that they're pretty okay at doing that kind of work right now. Smaller changes, like there are lots of things that, you know, we all have to do, like adding dark mode, for example, like smaller kinds of changes, where maybe being able to delegate those kinds of changes to a certainist agents are gonna, you know, get very, very good.

I'm excited about that. I think that the jury is still out on exactly what is the surface for being able to, you know, if you are the conductor of this orchestra, what is the right surface for being able to manage all of these things? And what is a realistic number of tasks for you to be managing at the same time? Because, you know, even I like a number of people have shown off like, hey, I've got like 20 different terminals open and I have quad code running in half of them and then Gemini running in, and it's, it looks, it's funny, but the reality is that you only have a finite amount of attention. And if you are going to be actually putting in diligence into your code reviews and into each of these, like workflows, you are probably only gonna be able to do a couple of things at the same time.

You know, it all comes back to multitasking as practices, but I'm interested in that, I'm interested in seeing where that goes. Another thing that's very interesting is, is vibe designing. I'm seeing that part of product, you know, product engineering, product development, evolving a little bit. It was very exciting to see, you know, Figma's MCP moving a little bit closer in this direction. Things that allow designers and developers to work more closely together, or to at least enable designers to be able to take their vision and turn it into a functional prototype, something perhaps a little bit closer to code that can then be productionized.

And it's not just a one-off demo. Yeah, I heard an engineering lead or design leader at Shopify said that her team, every designer on her team, so not all of Shopify has cursor and what they do is they create a Fitbit design and then they ask her to implement it and then they show it to engineers. And this is not saying ship it, it's just more, we now have something interactive that we can work with as opposed to just, well, they used to do what just sharing a Fitbit design. And I was like, this is something I've not heard before, like truly, like all designers or at least on a team using the developer tool, I can imagine designers using Visual Studio Code for, I mean, I can imagine them being able to open it, but I can imagine them, it's just not built for them. So this was really, really interesting to hear.

And again, this wasn't some vendor advertisement or anything. Yeah, I've shouted to the Shopify team. Honestly, I feel like a number of them are, have been very open to sharing what's been working out well for their teams. And I've enjoyed following their story. I think that more of these patterns spreading is gonna depend a little bit on training, on governance, on like clear boundaries between what is prototype code and production.

But already being able to turn something static into a semi-functional like prototype is very cool. Are we gonna see like everybody using cursor? I don't know. I think that I, at least for me, I see folks being able to accomplish the same outcomes from the tools that they spend the most time in, or the bridges between tools getting better over time. But I still think it's very, very exciting.

In the same way, you know, we're also seeing lots of good conversations about PM and EM roles changing. PMs maybe, you know, spending more time on problem framing and metrics and policies for agents, EM spending more time on emails and safety reviews and really enabling their teams to work with AI confidently. It won't change accountability for outcomes, but I am seeing a lot of good discussions about the need for taste in product and engineering, where that's gonna be what differentiates people. Because if anybody can, you know, look at what you've built and can use problems to accomplish, you know, a similar set of functionality, the taste piece is gonna continue to be very interesting for folks to focus on in the future.

You and I talked about sort of juniors and seniors. I think that, you know, there's gonna be interesting times for new grads versus seniors that, you know, AI definitely raises the floor, but it also raises the ceiling too. And juniors are gonna be able to get moving faster, but seniors who can, you know, write specs, decompose work, understand system architecture, review effectively, I think that they're gonna become even more valuable. And that last 30% that we talked about, I think it's leverage, not just busy work, but actually it's leverage. And, you know, a lot of surveys have been showing that trust is in a cautious, but optimistic place at the moment, but that cautious piece speaks to the need for that human oversight remaining pretty central.

Yeah, and if I think about, I was just thinking about, when it comes to this ideal parallel agent, I don't think it's ever been done in the sense that as a developer, you were never able to. I mean, we couldn't even kick off and talk with natural language to a machine that would spit out code that actually compiles. Like I think this is new as well, but the fact that you can do it with multiple. And one, it's not happened before and programming is a very, you know, you're in the flow, you work on one problem, when you stop working on it, you switch over, you get your context almost like the stack clears, you know, you load the new thing, it kind of feels like it, right? But on the other, when I think back to who are the best senior engineers I know, and whether they look like, well, they're on a team with a few mid-level, maybe some interns or new grads, and they're working on their stuff.

And then they're the ones who get the slack pinks and hey, could you please unblock me? So they context switch, review the code, you know, like criticize it, or they block out time and they review review review. And like for every kind of senior slash often these are tech leads, they have like a few people who are, it was not agents, they're people, but they just review their work and they kind of orchestrate them a little bit on the stand up, you know, they're in the planning meeting, they nudge them, they mentor them. So in some sense, I feel like we seniors already do this. And if I had a magic one saying, like who would I expect in a few years time to be able to manage multiple agents while the senior for sure?

Like new grads, probably, like, I mean, I feel that we stretch for them, but they're not expected to have them, you'll have the expertise. So I wonder if some of these skills are somewhat transferable in a sense that, and you know, that senior, why could that senior do that? Well, because the understood the code base, they knew what good look like, if they were always thorough on their reviews, they never let things live, they call out even the smallest thing. I completely agree with you. And I think that there's a lot to be said about how developer education in teams may also evolve for this moment.

Historic, I remember when I was coming up, you know, mentorship was always a big topic. And we talked about, you know, especially for folks who were joining a new team, the importance of like peer programming. I think we're gonna see, you know, perhaps even situations of trio programming, where it's a junior senior and the AI, and maybe the senior is going to be there asking you to explain the code that the AI is generating in some way, or walking you through how that code perhaps connects to other parts of the system. And really again, using it as an additional tool in their arsenal to be able to help build up confidence and awareness of how the full app actually works. We're also seeing some interesting discussion about, you know, potentially new roles or refinements of roles, things like forward deployed engineers.

I'm seeing interest in, you know, developers who are a little bit more embedded with customers who can build features rapidly with AI while feeding back requirements. And that type of thing might blur the lines between, you know, the developer, PM and designer roles a little bit. I'm interested to see where that may end up going. And I'm also interested in how, you know, just generally speaking, AI engineering is going to evolve how, you know, we approach education, whether you're in high school or you're in college. Like, are we gonna be teaching people prompt and context engineering best practices?

What does that all actually look like? How do we continue to enable people to think with systems design and engineering in mind? But I'm very excited in where the education side of this goes. So one area that we've talked about was code reviews and how it's really important, it's a bottleneck. And we should review the code.

One thing that I'm kind of noticing on myself already is when I use these AI tools, may that be cloud code or agents or even just auto complete, I have a tendency to do tap, tap or accept or just like accept all the things, especially when I'm working on something that, I mean, it's not the most critical thing in the world and it's kind of, it's easy. And especially after I started to trust that it gets it right most of the time in the end, I know I'm gonna review, but I find myself losing a little bit of criticality. I'm not as critical, you know, the second day and the third day. When I was the first time when I didn't trust this thing. And I worry a little bit that code reviews, the same could happen, you know, LGTM looks good to me.

I mean, I understand that's a Google has it built in as a feature in your code review tool, which is a really fun tool, as I understand. But there's this risk, there's always been this risk, of course, that like you just kind of like, it's a lot of work, it kind of looks good and you're not looking critically. How do you think, especially on, you know, like proper teams, we could battle or we could do this kind of fighting against like, yes, giving it a proper review, especially if we haven't written the code that much. Because I feel like writing the code, you have like two reviews. One is you're writing your own code, you're typing it out, and then someone else gives a review, knowing that it was you typed it out.

I mean, it's always more fun to write code than to read and review it. Yeah, a little bit, right? Yeah, and I think more and more of our job is going to become about reading and reviewing code. There are a few ideas that I've tried to suggest to folks. Things like, you know, for certain features or certain days of the week, maybe you intentionally try to not lean on using AI or NLM and just try to see like, okay, well, can I still solve some of these problems myself to preserve your critical thinking skills and force you to think about, okay, well, let's say that all of the top LLM providers were down for the day.

What would you do? Being cheeky, you know, I would probably say, yeah, I'm just going to go use Olama and a local model and I'll be totally fine. I'm going to find some fallback. But the reality is that I do think that it's going to continue to be very important for us to be able to think through how things work, be able to problem solve without necessarily, you know, relying on the AI. The models are going to continue getting better.

Their ability to take in sufficient context from your code base is going to continue getting better. But it is going to take quite some time, in my opinion, before you're going to be able to fully trust that in every situation, whatever requirements you throw out it is going to be able to get it right. And if you get stuck, if you're trying things out and you've tried it five, ten times and it still hasn't solved the problem, you're going to have to solve the problem yourself. So I think that being able to force yourself into situations where you're testing and retesting your critical thinking skills are going to be important. I also think that there's a little bit of value in, you know, doing some game theory around this.

Teams or individual people are probably going to start leaning on AI to do more of the code reviews, to keep up with the, you know, pace and velocity of change coming in. What are those workflows going to look like? If you have an agent that's saying like, yeah, I reviewed this PR and it looks good to merge in. Are you actually going to trust it or are you going to go and do a human level review, even if it take, you know, is maybe more shallow than he historically would have done or maybe you're okay with spending 10, 20 minutes or something on the roof. What does that work will look like?

In my case, even if an agent, an agent telling me that something looks good to merge is a signal. It's similar to a signal of like, okay, well, maybe a junior person on my team is also, you know, given this in LGTM, I'm still going to probably go back if it's critical enough. Or the CICD passes all the tests, right? It's greening, all test pass. Exactly, exactly.

It's a signal, signals are useful, but you still want to apply, you know, your lens on quality to it and take a look through and just make sure that you're confident. Things like having those tests, having, you know, whatever quality gates that you and your team discuss, like having those in place, all of these are good signals for building up confidence. So the more signals that you have to build up confidence, that things aren't going to go off the rails when you merge in. I think that's good, but I try to, you know, be very intentional with making sure that I'm not leaning on AI for absolutely everything. I may use it for many aspects of my life on a day, including AI coding, but I still try to make sure that I'm, you know, whether it's 20 or 30% or whatever amount of my tasks don't necessarily require AI, I still try to make sure that I'm using my brain to solve those problems myself.

I think that intentionally, like just be proactive with maintaining your critical thinking skills. I think that's going to help people. Yeah, and one of these things of like, you're still hands on two things, saying to my mind, one was a cloud code actually recently shipped. You can change modes, you can have exponents or any mode where it explains things. And you have a thing called learning mode, which is it pauses and says, you do this part, which I thought was really clever.

I actually turned it on for something that I was building. It didn't work as I expected because it gave me a really weird task to do, but I see the potential. And I actually, I actually unscented to lean a bit more on that and I think it's, I hope other tools will also do this part of the developer. I feel I do want to know that I can do it. I can do it, but how am I going to know if I'm not doing it?

Of course, you've gone around a little bit with it. I think that is such a fantastic idea. And if somebody is a junior or you're coming into a code base that you're not familiar with, quite that feeling of, the first thing I'm going to do is try to provide value and I'm going to just prompt a new feature into existence, maybe use the LLM to just explain how the code base works and to spend time soaking yourself up in the richness of, what is the code base? How does it work? How does everything connect together?

Before you start prompting things into existence. I think that using it as a learning tool is a very powerful thing. I mean, meaning rough. And this is also something that you said, new new new joiners using your learning tool. I'm hearing from a few companies that they're seeing new joiners on board faster.

The ones where they can use these AI tools to explain stuff to them out like the code base. It's just a kind of trusted person. 24-7, you can ask questions on like a senior engineer who, I mean, you can ask questions anytime, but obviously you don't want to do it. Like eight hours of the eight hour working day. Like you're going to keep your limits.

And it's interesting how that is changing the dynamics at a bunch of workplaces that do have these tools. Absolutely. I think that, you know, my hope is that it's going to become much more standard thing, training AI as a learning tool. And it's not just for understanding, you know, a new code base. It can be very useful for understanding programming concepts or frameworks or architectural patterns.

There are times when I want to be able to bring a feature over from one code base that is very different or written in a different programming language over to another. And I have found AI very much indispensable for those situations of helping me understand one code base and a path to being able to bring over something to a new one. And so encouraging your team to experiment with AI tools, sharing their best practices, making it a regular thing that people know it's okay to use it as a learning tool. I think it's just going to be great for you as leads and managers. And you're also a manager of a team.

And as part of this, obviously, you do performance reviews, promotions, help people basically, professionally, get better. How do you think the kind of your definition or what you're seeing on a team, a definition of a really standout solid software engineer has changed in the last three to four years since these tools have come out? What is new and what is the same? I think that what is new is the importance, you know, one of the things that I have found has always held true is the importance of being a lifelong learner. That has always remained true.

It doesn't matter if frameworks come and go, tools change, you know, the industry evolves, being a lifelong learner and being very open to trying out new things, failing, building up those skills. I think that that remains very, very important. The people on my team that I think were the most successful early on at leveraging AI for coding and for product engineering were the ones that went in with this mindset of, I'm happy to learn new things to try out and to have this growth mindset. And if it doesn't work, that's fine. At least I've tried it out.

At least I understand the constraints. And maybe I'll try out different models for these different use cases in the future. I think that has been very much a consistent thing that's been important. I feel like if you are a lead, now is the time for you to be helping your team through this moment in terms of showing that you're OK with learning as well. One of the things that I do every week actually is, I spend a lot of time reading.

I spend a lot of time reading. I will, of course, read your newsletter a lot. I will read a number of different papers, white papers, blogs, watch videos, watch courses, you know, read announcements about what is new in AI, what is new in AI engineering during the week. And I will surface those things to my team in a newsletter. I will also surface a lot of internal newsletter for you.

Yeah, I've got an internal newsletter. I write it every Monday. So I'll be writing it after this. And I will include, like, what am I hacking on? What am I writing?

What am I thinking? What are the things that I think are important for our team to actually pay attention to? And in this time where, you see so many people struggling to stay up to date, if you follow along on Twitter or other social networks, it can feel very overwhelming. Overwhelming, yeah. Because every few hours, it feels like something has changed in a fundamental way.

And being able to sift through that and help people to just pay attention to what is actually important. I think that's a great thing for leadership to be doing right now, especially if you can guide folks towards, hey, maybe spend a little bit more time poking at this rather than that. I just want to say, I think it's not done, because you're also staying closer to, you know, the industry, like, to being technical, maybe it's just right to say hands on. But you're pretty close to that just by keeping up with all of this. Yeah, yeah.

And I've, you know, I'll defer to people on my team to say things. But I think that I've generally had good feedback about this. I've even had, like, other execs saying, hey, actually, I'm finding it really hard to stay on top of this. And your newsletter is helping me navigate this moment as well. And so I think that as a lead being able to stay on top of what is happening at whatever level you're comfortable with or that your bandwidth allows, I think that that can be a really powerful thing for your team.

And it will help you just keep your own skills relatively sharp. We've been talking a lot about AI assistant engineering. If your team do have to build out, like, AI features, a lot of this stuff is also going to continue to be relevant for them, too, because you can help connect the dots between what is happening in the industry, where AI is concerned, that is actually important to their work versus what is not. And that's especially, I think it's exceptionally important in this moment where maybe there are things that they have not had to historically think about, like, oh, hey, well, Model X is really good at image generation. But Model Y is really good at generating sound.

I'm just making things up. But I think that it is really useful for you to be able to, maybe highlight some of these opportunities in a more concrete way. And then your team can go off and spend time getting into it themselves. But they're not starting from a blank slate. Plus, I guess you're just showing that, look, look, I'm learning, I'm spending time reading, understanding, digesting, I'm doing some stuff on the side, which kind of gives the permission that it's okay to do it.

And if anything, I mean, everyone for themselves, but I would assume that right now, because we are having a big change in this technology, it's a new tool, a lot of capabilities. It should be okay at most companies, and most scenes, unless you're like, has done crunch time shipping something like on Friday, to for engineers to spend some time experimenting, figuring out, hey, can I use this thing? Will it work for me? When I talk with again the folks that shopify, they do a lot of this. And in the end, actually, what some people think might happen is like, well, people do less work.

Actually, what happens is they do the same work as they did before, but they were motivated. They tried a bunch of new stuff. Someone that doesn't work, and it's kind of a, you would think it's a waste, but it's actually not. It was learning. And now they're a lot more confident about what works, what doesn't, they'll confidently say, like, okay, we're not gonna use AI first, generate this part of the code base for prototyping, it's great, and so on.

So I feel like everyone's figuring things out. So it's almost a waste to non-gift permission. And how better ways to get permission than to show that you're doing this as well? Yes, yes, absolutely. AI is a tool, tools that can compress work, I think can also lead to good moments of clarity, and it is faster than ever to try out certain classes of ideas, and I think that that efficiency can also sometimes highlight the importance of our human qualities when it comes to judging.

What should we be moving forward with? What should we be ignoring? What should we be minimizing? I think that trying to embrace AI in this moment, if you're at what, you know, what we're doing for your union teams is a great use of time. I did wanna say, if there are folks in your audience who are in enterprise, especially, I know that having gone through this journey, there may be teams who feel like, oh man, it feels like startups are able to move, so much faster than we are, or they're able to use all of these great new tools and these models, but we are still waiting on the enterprise-friendly version of these things, and I've talked to teams where they can sometimes feel like we're waiting, we're waiting, we're waiting to be able to try out more of these tools.

What I ended up doing in my team was, we were similarly, a couple of years ago, now we were similarly waiting for the company, embrace the official way of doing things, but that didn't stop us from being able to learn. So I encourage people, well, a lot of us are hacking on side projects, right, at the weekends. We can try out whatever third-party tools, third-party models, we can try out and learn, and that doesn't have to be a big blocker for you. So you can still help your team along this journey without necessarily having to do a lot of that waiting. So just one of the left folks know, there are many paths to be able to embrace this moment.

Yeah, I think the consensus is pretty much that these tools will be widespread, they keep changing and they're already contributing to a lot of teams, code-based and people using it, so it might as well just use it, and it takes time in the end. So the earlier you start, you'll have the only learning to do regardless. Yes, absolutely right. So as closing, I just have a few rapid questions, so I'll just fire them and then you tell me what comes to mind. What's your favorite programming language and why?

My favorite programming language is JavaScript. I thought you might say this, but I wasn't sure. Of course, you wrote books about this. My favorite programming language is JavaScript, and it's not for the personal reasons. There are probably better programming languages.

I like JavaScript because it enables anybody on the planet to be able to build and ship something to the web in a way that doesn't require gatekeepers, it's very open. And I think that there's something very liberating about that idea. So that is one of the reasons why I like JavaScript, but probably the most. I like it. I'm always surprised when a software engineer says JavaScript because we know that compared to languages, it has a lot of limitations, and there's books written about it as well, but I cannot disagree with it.

I love it. It is true. What's a tool that you really like using? And what does it do? I would say that's one of my favorite tools at the moment, and I'm biased.

I have a book out about this as well. It's probably Bolt, Bolt.new. Bolt.new is one of those vibe coding scaffolding tools that people can use. They very recently added support actually for being able to use custom agents. So you can use Cloud Code, for example, to build your vibe coded app.

And the output is generally very high quality, great design. So I've been really enjoying that. And the team has been, I would say, on the edge of offering some really great integration. I've been liking this idea that many vibe coding platforms are now starting to think about the integration layer. So how can we automate the idea of you needing to use, you know, like a super base or an authentication provider or these other things?

You still need to pay, of course, attention to all the code that's being generated, but being able to remove more of that set of friction, I think is amazing. And I guess the fun fact on Bolt.new is they started off as a company called StackBlitz, who built a really cool online editor, really advanced, and then they moved over from that to Bolt.new. So their software engineers behind it, again, who really know their craft. Yes, yes, absolutely. And finally, what's a book that you would recommend?

So the book I'm going to recommend is one that I think covers, I would say career paths and best practices for software engineers in a way that I found very compelling. It is by you, it's the software engineers guidebook. And of course, I'm playing the crowd, but it is generally a very, very sharp book. From all the reviews that I've read about it, other people have also found it to be an excellent read. Well, it's not my best, but we didn't talk about this, by the way, so this recommendation came just as a surprise to me as a teller.

Thank you. Generally an excellent book. I guess if I wasn't going to, of course, record without one, I think that if you are trying to learn more about the foundational aspects of AI engineering in this moment, there's a great book called AI Engineering by Chip Huyan that is also, we're taking a look at, very, very well reviewed, a very, very thorough book, and I also recommend folks check that out. It's such a good book. I feel if you know the concepts there, you're pretty well off and if you don't know them, you're probably going to be still finding your way.

So big plus one for that. Addy, this has been great. Thanks so much for doing this. And this really awesome conversation. Well, thank you.

I really appreciate it. I hope the folks, if you are interested in the space beyond Vive coding is out now from O'Reilly, hopefully it'll be useful to some people, but it's been a pleasure having this conversation with you. Yeah, and I've been reading the book. I really like how it goes into practicality. So I can also very much recommend it.

Thank you. I hope you enjoyed going beyond Vive coding, pun intended, with Addy Awesmani. One of the things I took away from this conversation is how it's really important that as an engineer, we keep understanding exactly what the LM does and why. And if we don't understand it, we just stop, understand it, and then move on. As long as you understand the LM, you are in charge.

But once you stop understanding, it's in charge, and you can kind of become helpless and out of your depth. For more tips on how to get up to speed with AI engineering, check out DeepLabs in the Primatic Engineer that I wrote. We'll share a link to the show notes below. If you enjoyed this podcast, please do subscribe on your favorite podcast platform and on YouTube. A special thank you if you also leave a rating for the show.

Thanks and see you in the next one.