1045: Who can actually max out 5 AI subscriptions?

Scott and Wes answer listener questions about what replaced blogging, what to do while your AI agent works, and who should pay for token costs. They also cover keeping agents from going off the rails, getting your team excited about AI, and staying motivated when AI takes the fun out of coding. Show Notes 00:00 Intro 00:40 Welcome to Syntax! 01:37 What’s the Modern Day Equivalent of Blogging? 08:35 What Should You Be Doing While Your Agent Works? Vibe Stretch 13:23 Brought to you by Sentry! 14:02 How Much Should Your Employer Cover for Token Spend? 18:32 Tracking What You Actually Burn with Codex Bar 20:26 How Do You Code HTML Emails in 2026? 25:22 How Do You Keep an Agent From Going Off the Rails? Swamp Club 30:23 Steering an Agent Mid Turn: Which Harness Does It Best? 32:44 How Do I Get My Engineers Interested in AI? 35:48 Beyond Telegram: Better Interfaces for Non Coding Agents 38:41 Staying Motivated When AI Takes the Fun Parts of Coding 41:55 Killing Rogue Node Processes and Restarting Your Machine 44:12 Custom Internal Harnesses: Worth It or a Token Hog? 47:47 Sick Picks + Shameless Plugs Sick Picks Scott: Driving Rogue Wes: Amazon Basics Electric Screwdriver Hit us up on Socials! Syntax: X Instagram Tiktok LinkedIn Threads Wes: X Instagram Tiktok LinkedIn Threads Scott: X Instagram Tiktok LinkedIn Threads Randy: X Instagram YouTube Threads

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Scott Tolinski?

  It's crazy to me to hear people be like, I got five different subscriptions and I maxed them all out. Like, that feels performative to me. I mean, I can understand maxing out one or even two, honestly, but maxing out more than that feels crazy. Like, what are you doing?

• 0:11Wes Bos?

  Yeah, there's people have like 20. I see this all the time on Twitter, but like, I have 20 different subs and I've maxed them all out. Like, what are you building?

• 0:19Scott Tolinski?

  It's cheaper than not paying for an employee. It's like, well, okay, but what are you, what do you need?

• 0:23Wes Bos?

  But what? Show us. Show us the amazing things that you're building. I don't see anything other than you're yapping on Twitter.

• 0:39Wes Bos?

  Welcome to Syntax. We've got a potluck episode for you today. What should you be doing while your agent is working? How do you stand out? Is blogging dead? How do you keep your agent on the rails? How much should your employer cover for token spend? How do I get my engineers interested in using AI? And finally, a non-AI question, what are you using for emails? If you have a question, make sure you send it to us. Go to syntax.fm and click on the potluck button. Submit your question and we will answer it in an upcoming episode. Also, we want to tell you, San Francisco, we are coming to do a syntax meetup on October 27 at Bear Bottle Brewing. So if you are anywhere near San Francisco, or you want to, dare I say, hop on a plane, train, automobile,

• 1:23Wes Bos?

  and come down to San Francisco to hang out with us, we are going to be hosting a meetup. October 27, go to syntax.fm/meetup to grab your ticket. It's free, but you do need a ticket to get in. We'll see you there. First question from Alex says, what's the modern day equivalent to blogging? I thought this was a really cool question. A few years ago, writing about technical things was a way to get noticed, showcase your skills, and maybe land a job as a developer or product designer. I feel like people don't read blogs as much anymore. What has taken its place? TikTok, Instagram videos, and what do you actually talk about? Are people still interested in seeing someone share their technical knowledge? I'm a product designer and I used to build front-end prototypes. These days, I mostly work with Claude to make them.

• 2:04Wes Bos?

  I really enjoy the field and what I do, and I'm not looking for a way out, but I don't want to get lost in a sea of people doing the exact same thing. I want to stay relevant. So, long story short, how do we promote ourselves? Now, this is a great question because we even saw when, when Jeff got announced, there was a couple of people being like, we, we built this and it just didn't have the like hilarious Twitter video to, to get posted with. And as frustrating as it is, it's not enough to simply just like build something and put it out there or share everything. That used to be it. That was, that was awesome, right?

• 2:45Wes Bos?

  You, you work really hard on something, you learn a lot about a specific topic, and then you can write a blog post and people will love to hear about that. And now there is just 10,000 times the amount of content out there. That's hard to just get people's attention. And I think like, there's also, a lot of this content is really good and people like to watch stuff that is well-presented and entertaining and whatnot. So, how do you, sorry, I was going to say that again. So, how do you promote yourself right now? Right now, it seems to be that people are looking for opinions, right? There's just so much stuff out there and people are looking for somebody who has taken the time to figure something out

• 3:28Wes Bos?

  and has garnered some opinion on, on these new pieces of technology. They have experience with it. Maybe they've done some exploration. Often, the stuff that is most interesting to me is somebody who has been interested in a specific aspect and they have not just done the like high-level six prompts and, and make a demo. They've actually gone into it and used their noggin about how should we best approach this thing? What, what might it look like if we were to design with our voice, you know? There's all these like people that are on the next level of like, how do we use this stuff

• 4:12Wes Bos?

  to do something that is not just build the same stuff that we've always done, just faster?

• 4:18Wes Bos?

  And, and for those people who are able to distill their thoughts cohesively and, and do it in a interesting manner, I think that those are the people that can promote themselves. The mediums, I still think like blog posts are still a great way. You're not going to get the crazy like million views, but like Hacker News is still mostly just text. And, and those are, are the things that are landing on, on Hacker News. Obviously, posting to social media is very good as well.

• 4:50Wes Bos?

  The like short-form video is its own thing altogether as well, but it's certainly not a bad way to practice being able to say, I am passionate about this thing. I learned about X, Y, and Z. I have opinions on this thing. Figuring out how to distill that down into a slightly entertaining 60-second video, it's, it's not, not a bad skill to have in, in going forward.

• 5:19Scott Tolinski?

  Yeah. I think in general, yapping is the new blog.

• 5:23Wes Bos?

  Yeah.

• 5:23Scott Tolinski?

  Blogging, yeah. Although, you know, there's a lot of yappers out there and I think the thing is, is that like, people, there's always, as you mentioned, room for somebody who knows what they're talking about and communicate it well. Where what we end up getting instead is like 1,000 awful YouTube channels being like, this changes everything with a bunch of just bad or dumb opinions or not informed, just racing to get something out, which is always frustrating to see, right? When you, I remember being really frustrated trying to find any good information about Hermes because every single YouTube video about Hermes with some AI bro who doesn't know his,

• 6:07Scott Tolinski?

  I'm going to say his ass from his elbow, is that a, is that a saying?

• 6:10Wes Bos?

  That's it. You got it.

• 6:11Scott Tolinski?

  That's, yes, got it. I've never, I've never said that before, who just trying to get clicks and views and whatever. So, you know, there's always going to be room for people who know what they're, they're talking about.

• 6:22Wes Bos?

  Yeah.

• 6:22Scott Tolinski?

  And who can do it well. Yeah.

• 6:24Wes Bos?

  It used to be also like, we would just say like, hey, build stuff and push it to your GitHub because when people go looking to hire somebody, they're going to see that you have stuff on your GitHub and you know what you're talking about. And it's not just the same to-do list that you, everybody did when they were learning rails. And that's a little bit different now because everybody can push a whole bunch of slop to their GitHub and it's really hard to go through that. But if, if, if I was looking to hire somebody and they had some like neatly worded blog posts or just like little investigations into, like, I'm following this guy right now on Twitter. He works for Netflix and he's rebuilding the Micro Duck right now and the Hugging Face Micro Duck.

• 7:03Wes Bos?

  And he's just sharing little bits of information about what he's learned, little things that he's run into. And it's just, just photos and just being like, oh, I had, had trouble with X, Y, and Z. And like, if I was looking to hire a really smart guy, that guy would probably be at the top of my list right now. And, and he's not doing like some influencer social media. I think that's a totally different game. He's just like sharing interesting stuff and able to communicate it in, in a neat way. In this case, it's just texting photos on Twitter. So.

• 7:36Scott Tolinski?

  And that's a lot harder to AI fake, like.

• 7:40Wes Bos?

  Totally.

• 7:41Scott Tolinski?

  YouTube videos, easy to AI fake. Blog posts are like, who's, who's actually writing good blog posts anymore by hand, right? Like, the blog posts that I read, I take one look at them. I read a paragraph and I'm like, all right, I'm done. This is like, this is junk. So, yeah, yeah, I agree. That kind of stuff that is connectable, that is human, that is relatable, that is, you know, real.

• 8:05Wes Bos?

  Exactly, exactly. So, just, I think we could sum that up with just like, do interesting stuff, put it out there for people to, to see. It's not going to like go super viral, but I think if, if it is that you're, you're trying to stay relevant and have opinions on all of the stuff that is changing right now, I think that that will stand out significantly when it comes to employer looking at, does this person know what they're talking about?

• 8:33Scott Tolinski?

  Totally. All right. Next one here from Elias. While an AI agent runs for minutes at a time, do you two actually get up or do you sit and watch it?

• 8:47Scott Tolinski?

  I write with Claude code all day and I noticed I never move during those waits. So I wrote a hook that prints one stretch line into the session while an agent is still working. It's not a timer. It counts only the minutes and agent turn kept me waiting. So, a day of quick turns never fires it and stepping away for 45 minutes clears the count. You both take training seriously. Yes, we do. So I'm curious what you do instead of total time. Tool, the tool is open source. If anyone found it useful and they have a link to it called Vibe Stretch. I love stretching during working. What I do typically while one agent is working

• 9:30Scott Tolinski?

  is I put another agent on a different task and then I have another agent on another task. And then I have another agent on another task. And by the time one finishes, another one has started and you just kind of go into a loop until your brain fries. But yes, what, what should you be doing? Man, it would be like, it would probably be good to do pushups or get the blood moving or go for a walk or, yeah.

• 9:54Wes Bos?

  Yeah. Man, like, this is a great question. Like, what should you be doing while your agent is, is churning?

• 10:00Scott Tolinski?

  And I will, well, before you get into that, Wes, I will say, I don't have agents running for that long typically. If an agent for me is running for like 10 minutes, I'm getting kind of annoyed. I like to have a tighter loop than that and that's a mean thing. Yeah.

• 10:15Wes Bos?

  But even, even if you're waiting for like two or three minutes, that's enough to sort of break your context. And that's what everyone's talking about, their flow getting, like, you don't get into that flow state anymore. What I've been trying to do is rather than just go like, go on Twitter or whatever, is just like, take a look at what it is that you're building and, and think about what could possibly done. There's so much, specifically in the UI of your applications where you can tell an agent, hey, we're building X, Y, and Z, but there's so much more time needs to be spent on like writing down and detailing what needs to be fixed with your UI application. So, what I was doing last week is that.

• 11:00Wes Bos?

  I set up one orchestrator agent. I set up another one, which I just call my GitHub issue one. And this is just a very cheap model that is running. And the orchestrator will just peg off GitHub. It'll take a GitHub issue, work on that one. And then as that's being worked, I'll go into the GitHub issue one and say, hey, I'll just sort of just like brain dump everything that is wrong with the application at the current state, you know? When I click on an item, I shouldn't be able to drag and drop it. You know, I was building a 3D gym builder. I'll show you a demo of it in a coming episode. But there was just like

• 11:45Wes Bos?

  so much around what the UI looks like and all of these little things. And quite honestly, it took longer for me to like, figure out these things, think how we should better approach this, research, what are the design patterns around this, and then distill that down into a GitHub issue. It takes longer than to do that than it is to like, have a single turn in it. So as the agent is churning, try and figure, take a look at your application. And whether it's what you're working on right now, or whether it's something else unrelated, and try to document in detail all of the problems that you have, so that when it does come time to pick up that issue, you have this huge backlog of information, rather

• 12:29Wes Bos?

  than just you just saying like, make the UI better, more modern, you know? Do a bit of research.

• 12:35Scott Tolinski?

  Yeah, yeah. People who don't do a lot of AI coding, or they use it in still like an autocomplete type of way, might think like, that's crazy, this conversation in general. Like, oh, I see every line that it writes, or whatever. But the way that we write software with AI is different than how you wrote it before. Because again, you are chiseling away at a statue that is gotten close the first time, and then you're refining, and then you're refining, and you're refining, rather than building something up from the ground up, right? So that process requires that eye and that intent to refine, right? So I like your point about being able to spend that time, like, using that time to understand

• 13:20Scott Tolinski?

  what needs refining and to review and those types of things. And if you want to see all of the errors in your application, you'll want to check out Sentry. At sentry.io/syntax, you can sign up today and get two months for free. Sentry is just a really incredible tool for not only tracking your performance, making sure your application has no bugs, but even just seeing what goes wrong when something goes wrong, because things go wrong all the time when we're coding. And you don't want a production application out there that, well, you have no visibility into in case something is blowing up, and you might not even know it. So head on over to sentry.io/syntax. Again, we've been using this tool for a long time, and it totally rules. All right.

• 14:02Wes Bos?

  Anonymous says, how do companies handle employee AI usage? An employee at home, working on their own dime, may only do a couple prompts and expect an outcome. However, the same employee at a work with near infinite company budget may prompt and prompt and prompt until they do something that feel is close to perfect. I feel this would cost the company a boatload of money. Do companies give each employee an AI budget? Great thing. And I think we have seen in the last six months, a lot of companies changed their stance on this.

• 14:34Scott Tolinski?

  Yes.

• 14:34Wes Bos?

  It was initially, everyone's like, doesn't matter, like, just like, this is a magic box, we can do everything. And then the spend got a lot out of control. And like, the slop cannons that people were shooting off started getting overwhelming. And the problem starts to be how much it costs, and how much, like, somebody has to review all this code and security issues, all of that type of stuff. So I did ask on Twitter, how much is your employer cover for AI usage? Do you get subs? Budget unlimited? Handful of people unlimited. Couple people said they will give us $200 for a sub. Some people say they, $500. A lot of people that are on like, enterprise. The enterprise for a lot of these big things,

• 15:18Wes Bos?

  Anthropic and Claude, they build based on token. So it's not as good of a deal as the like, $200 a month sub. Some people said they get several subs. So people get covered like, $500 a month, and they'll use that to buy a Codex and a Claude sub. And they'll use one and run that out, and then use the next one. Seems to be all over the place, but the overwhelming majority of them seems to be around the $200 to $500 a month thing. And I think this is something that people have to, at least right now, people have to learn to use, which is like, you have these resources available, you can't just willy-nilly

• 15:58Wes Bos?

  run up. Like, you could very easily spend like, $1,000 a day as a dev, if you are building lots of new features and constantly scanning everything. So it's a bit of a skill as to like, how to appropriately budget this type of thing. Whether this will be a problem in six months from now, has to be seen. But for right now, I think most people are getting like, a 600 bucks, 500 bucks, 200 bucks a month. And then the employee is, it's left up to the employee to figure things out.

• 16:29Scott Tolinski?

  Yeah, because what, you just rip with API token based anything, and next thing you know, the HR is using Vable 5.1 to draft an email, and just spend is getting crazy, you know? I mean,

• 16:43Wes Bos?

  People don't realize how much it costs. Or you set up like, a reminder, and like, to check your email every two minutes, and like, all of a sudden, that's $10,000 a day?

• 16:54Scott Tolinski?

  Yeah, normies do not understand the cost. They're like, oh, I got a ChatGPT subscription at home, or whatever. It just works, and it's done, right?

• 17:03Wes Bos?

  Mm hmm.

• 17:04Scott Tolinski?

  But yeah, when you're going token based, it's crazy.

• 17:06Wes Bos?

  I think it's up to you to understand also, like, it's good for you to learn how much this stuff costs. And if I do choose a cheaper model, what does the output look like? You know, like, just like, we had to learn how our performance works in our applications, and how to write code efficiently. I think, at least right now, a skill you need to have is how do I take these resources and get the most out of it? Because certainly, a lot of these people with these, I got a couple DMs from people with like, $15,000 a month, they get access to this type of thing. And you can go absolutely nuts. And if you have like, a leaderboard inside of your company showing who's spent the most tokens, people are just going to be churning these and creating all sites with unnecessary features. So I

• 17:51Wes Bos?

  think there's some restraint that can be had in these things. And a good way to practice that restraint is by a somewhat of a limited budget, with the ability to then scale up if you need to, you know? Like, if you're like, hey, boss, we're rewriting in Rust, I need 20 grand of tokens, you know? Like, that could possibly be a way to get additional help. How would you even?

• 18:15Scott Tolinski?

  Like, a token approval, like, approval token budget for a specific thing, rather than just yellowing all the time, right?

• 18:22Wes Bos?

  Yeah.

• 18:23Scott Tolinski?

  Yeah, I'm sure we'll see way more of that, as companies try to figure out what the heck to even do. Yeah.

• 18:28Wes Bos?

  I will say, though, that the $200 a month subs is by far the best. If you look at how many tokens you use.

• 18:36Scott Tolinski?

  Easily the best, yeah.

• 18:37Wes Bos?

  If you take a look at like, the token, I use this app on my Mac called Codex Bar, and I hook up all of my subscriptions to it, right?

• 18:47Scott Tolinski?

  It's just, it's called Codex Bar, but it works with all of them?

• 18:50Wes Bos?

  Yeah, it's from the like, Claude Bot guy. But it has like, Codex, I've got Codex, Claude, Cursor, Gemini, Copilot, and ZAI, right? Those are all the subs I have right now. And it says, in the last week, I've used $1,100 worth of tokens. And that's not even including stuff that's not on other computers or in remote sandboxes as well, right? And I certainly have not paid $1,100. So those, these $200 subs you can get from, these are by far the best deal right now.

• 19:25Scott Tolinski?

  Yes, absolutely, totally. If your company's paying for one of those, get it. That's what I'm gonna say about that. Because they do last, it's crazy to me to hear people be like, I got five different subscriptions, and I maxed them all out. Like, that feels performative to me, like, because, I mean, I can understand maxing out one, or even two, honestly, but maxing out more than that feels crazy.

• 19:46Wes Bos?

  Yeah, people have like, 20. I see this all the time on Twitter, but like, I have 20 different subs, and I've maxed them all out. Like, what are you building?

• 19:54Scott Tolinski?

  It's cheaper than not paying for an employee. It's like, okay, but what are you, what do you need?

• 19:58Wes Bos?

  But what? Show us. Show us the amazing things that you're building. I don't see anything other than you're yapping on Twitter.

• 20:05Scott Tolinski?

  Yeah, right. Yeah, that's what they're building. They're building a Twitter yap farm. That's right.

• 20:10Wes Bos?

  Oh, maybe the app that you're working on, where you like, show all of your projects? Like, before you delete a project from your computer, it should show you how many tokens you spent on it. And you're like, this is a dumb idea. Yeah.

• 20:24Scott Tolinski?

  Are you sure you want to remove $54 worth of shame here?

• 20:30Scott Tolinski?

  Okay, next question from Cosman. How do you code the emails in your application sends in 2026? Like password reset, welcome ones. How do you just go with whatever your? Do you go with whatever your framework provides? Or do you go down the rabbit hole of rolling your own HTML emails? So I've always rolled my own unless I'm using something like PocketBase that has things built into it. You know, I don't feel like this, especially in the age of agents. This becomes very easy to do. But transitional or transactional email is one of those things I hate paying for. It feels way more expensive than it should be to have an app that sends an email once in every while.

• 21:12Scott Tolinski?

  I've been using resend lately, but I do find it's like, I don't know why I'm paying so much for it. How do you feel about that? What do you use, Wes?

• 21:20Wes Bos?

  I use React render to string. And it works great because like you just like emails are not that complicated. The support in them is pretty good. And all you need is some sort of templating engine where you can make reusable components. Right? React is really good for that. And then you need some way to render that to HTML, which all of these frameworks have. And then you also need some way to like inline your CSS because that's how you get the best thing. So you take your CSS and it will figure out what applies to what and you can put it in there. I know there are lots of other things out there. Like I know it was React email from the resend folks as well.

• 22:05Wes Bos?

  I have never used resend, mostly because I don't want to use something that is specific to one company, you know? Like a lot of these solutions that people praise.

• 22:16Scott Tolinski?

  I use resend for sending. I don't use their React stuff.

• 22:19Wes Bos?

  Okay, do you use SMTP or do you use their APIs?

• 22:24Scott Tolinski?

  I use their APIs. Yeah, you're right. Yeah, that is still using their stuff. Yeah.

• 22:29Wes Bos?

  Yeah, like just use SMTP and then use any SMTP, you know? SendGrid is probably a good one that a lot of people recommend.

• 22:40Scott Tolinski?

  But they're still so too expensive.

• 22:42Wes Bos?

  What's too expensive? Sending emails?

• 22:44Scott Tolinski?

  Yeah, yeah. How much does SendGrid cost? Let's see.

• 22:47Wes Bos?

  You know, I was using, or I am, what am I using?

• 22:51Scott Tolinski?

  Twilio now? Twilio owns SendGrid now? I miss, what was I using? Mailchimp's one, Mandrel. I was using that for a while, but even then it's like not that cheap. I like want something that's dirt cheap. That's what I want. I want something that's free, something that I don't have to pay for.

• 23:06Wes Bos?

  You can't. That's how you get it thrown into the spam box. I'm on, I'm still on Postmark. I have been for like 10 years. They switched recently from buying credits, which I loved because if you send a whole bunch of email at once, you just pay for that. And then if you, if you have like a little bit of a lull in between sending them, then you're not paying for anything. And software as a service doesn't like that because they like you to pay a monthly fee and get a certain amount of sends and all this BS. So I'm still on Postmark just because of the pain in the butt that it is to set up sending emails. But I know that I could switch.

• 23:42Scott Tolinski?

  It does seem like it is free for 100 emails a month. So side projects with low volume, which is something that I'm often needing email for.

• 23:54Wes Bos?

  Yeah, I would probably say go get hooked up with Amazon AWS or what is it? SES? Because then you just pay by the email. And then if you do need some of these other features like open tracking and whatnot, you can, you can swap over pretty, pretty easily. But always, always, always go based on standard APIs. No matter how nice some of these features might be and whatever, it's so much better to, to use basic email. So if something happens or all of a sudden your emails start getting caught in spam box, you can move over to another provider.

• 24:32Scott Tolinski?

  Have you used CloudFlare's send emails?

• 24:36Wes Bos?

  No, I have not. Actually, that is something that I need to try out. It's one of their newer products.

• 24:43Scott Tolinski?

  Yeah, as we're talking about this, I'm thinking, wait, I did see that they do have email routing SMTP. You can use any node mailer or anything like.

• 24:53Wes Bos?

  Yep.

• 24:54Scott Tolinski?

  Yeah, okay, you know? Yeah, I'm listening to CloudFlare.

• 24:59Wes Bos?

  It's about deliverability. A lot of people are very sensitive to switching these types of things because all of a sudden 2% of your things are now getting caught in spam. Like that's a big cost because now those people are not buying your products. So people are, people spend a lot, a lot of money. That's why resend is such a big company. Email in people's inboxes is, you can make a lot of money doing that.

• 25:24Scott Tolinski?

  Word.

• 25:25Wes Bos?

  Miles Dyson says, I overhear interviews in my office and they're always asking candidates how to make sure that agents can't go off the rails. I'm not sure people asking the question know the answer themselves. Sign of the times. What do you guys think? How would you answer this in an interview? And can it actually be done reliably? I put this one in here because I know Scott has lots of opinions on making things deterministic or having rules and tools in place so that things can't go off the rails, right?

• 26:00Scott Tolinski?

  Yeah, yeah. So, you know, one thing that people don't realize is that their harness often has possibly things like hooks, like Claude Code has hooks that can deterministically step in if your agent is about to do something like RM, right? I mean, there's also deny lists and things like that to prevent your, your, your AI from doing stuff. But just to use a very simple example that everyone can understand for a hook, you can say, is this agent trying to do an RM? Then provide this response to the agent and block it. Say, you can't do this, do this instead. And that works for any type of time you want to steer an agent to do something when it's about to do something else.

• 26:45Scott Tolinski?

  As far as controlling them though, there is only so much you can do.

• 26:51Wes Bos?

  Yeah, there's obviously agent rules and linters and whatnot, right?

• 26:57Scott Tolinski?

  There's, there's like, you want to put things out of the hands of the agent and say, okay, if you're doing this, go into this path. If you're doing this, go into this path. I use something called Swamp Club to deterministically route a bunch of things. And I really like that for that. That's not going to like keep you safe, but it is going to give the agent a guaranteed path to say, oh, like here's a perfect instance for why Swamp Club is great. The other day I had my Jellyfin go down and I wasn't sure why. So I go to my agent, hey, my Jellyfin's down. Can you check it out? And the agent goes off and does all this stuff and it spins up a new copy of Jellyfin on a, in the Docker that I didn't ask it to do, obviously. And that's the exact type of thing where it's like, it totally went off the rails.

• 27:42Scott Tolinski?

  Not what I asked for it to do. So instead what I did is I put in a swamp automation for checking the status of my existing Jellyfin, where it lives and all that stuff. So now in agencies always asking about Jellyfin, let me hit that automation. The automation returns a status. It then knows what to do, like a status and instructions and it knows what to do next. Right? And it takes it from a flow of being like, oh, let me, let me start from scratch to let me have a more guided pathway into doing things. So I use that quite a bit for all kinds of stuff that I want to be able to keep it on the rails more. Whether that is like orchestrating like steps, this, then this, then this. But there's still times when the agent's going to do whatever the hell it wants.

• 28:26Scott Tolinski?

  So that's why people like CJ have recommended putting things in isolation, putting containerizing and then running your agents in a container. They could only do so much harm there in some sense. And then like, it's not practical to me.

• 28:41Wes Bos?

  In addition, yeah. In addition to like, like going off the rails and causing harm, I think this question is also asking just like, how do we make sure that it's, it develops the things that we want and doesn't start like adding a whole bunch of stuff? And partially you can do that in like your agents.md, you know? Like don't write fallbacks for absolutely everything. That's one thing that happens all the time to me right now is you change something halfway through on your second prompt and it's like, well, we must support the original use case forever. And it just does six if statements in there everywhere. So like partially that can be done in your agents.md. Partially that can be done with deterministic tools where you like linters and tests.

• 29:25Wes Bos?

  Like one of them I wrote the other day, which was when I was working on some 3D stuff, which is a whole bunch of tests that will spin up an agent browser to check the frame rate of rendering. Because very easily you can just add features all day long, but that can really start to lag down the UI. And every time they implemented a feature, it would spin up an agent browser and then run the frames per second and make sure that we're still getting over like a 55 frames a second during that. And, and those are the types of things that you can do. But also if you find that it is going off the rails a little bit too much, that's, that's probably on you and you need to be a little bit closer to these types of things. And you're maybe just YOLOing the prompts a bit too much.

• 30:10Scott Tolinski?

  Yeah, I do, I do. I'll watch the output and be like, and I see it getting into paths that I doesn't need to be getting into. And I'll spam that escape button and be like, you asshole, what are you doing? I didn't tell you to do that.

• 30:23Wes Bos?

  Talking about the escape, this is totally, totally a different question, but just like, like steering in general. I use all of the harnesses and sometimes when you send like, no, don't do that, or you want to like tell it to do something, you have to hit like escape and just kill the whole thing. Some of them, like Codex will send like the command the next turn. So it'll like, it'll wait until it comes back and then grab your next message. Some of them will wait until you're totally done and then it sends it and it's got to kind of backtrack and you gotta hit X on it. I found that I like the Codex. I don't like the open code implementation. There really is no steering and I know people are like, you shouldn't do

• 31:08Wes Bos?

  steering at all, but there's so many times where I'm just like, I want to quickly add in a little bit of information. I know Claude also has like the by the way, where you can ask questions, or it will then tell the main harness to update something. Do you have any thoughts on the best steering approach?

• 31:24Scott Tolinski?

  I don't know. I don't think that's something that I figured out really well. I don't really love the by the way kind of features and stuff like that. There's a by the way feature in a lot of these that again, it's just like kind of adding additional steering. Sometimes that turn takes too long. Typically when I'm stepping in with some additional information, it's like urgent, otherwise I'm not stepping in. So I almost always default to stopping regardless. I think the one that I don't stop and then type is Claude. I do let that one pick it up. For some reason, I found that Claude code does a really nice job of that in particular, but I don't mind just hitting stop on that bad boy because it's like-

• 32:03Wes Bos?

  So many times of like, I just like, I don't want to stop it because it's working and I don't want to have to restart everything. And it's not important to stop it because it's not doing anything bad. But like, the next turn you have, which is probably in the next 20, 30 seconds, you should throw this message in there as well for some additional context. And maybe you want to change course or I guess steer yourself. So I've, I also have not been super happy with any of them, but I do think Codex is my favorite so far.

• 32:34Scott Tolinski?

  Yeah. Folks, what are you doing for that? Are you using by the way? Are you just steering and waiting for it to pick it up? What do you like to do to keep that going? Because I'm looking to optimize here. I'm interested in that. Yes. All right. Next question from Ash. I'm the head of an engineering at a small startup. General interest in agentic development is on the lower end from my engineers. So it's mostly me experimenting and trying to get folks to try things. How can I get my engineers more interested on their own? Would love a highly engaged culture around this stuff. Yes, Ash, you know what? It's going to be tough for some people because they have like an ideological like anger towards integrating

• 33:17Scott Tolinski?

  this stuff, right? They don't want to do it. But I think a lot of people too, who have put up a huge wall into agentic coding, like I'm not going to do it. It's not me, whatever. I don't think they've realized the progression of the models and how good they've gotten.

• 33:37Scott Tolinski?

  So I don't know, like, is a demo, look at what my agent can do going to work? Because I think a lot of these people have-

• 33:45Wes Bos?

  I think those like, look, one prompt things hurt more than they actually help.

• 33:50Scott Tolinski?

  Right. They hurt. Right. So like, having someone shadow you, I really don't know. If you're the head of engineering at a small startup, to me, it seems like, like more lunch and learn type of deals where you can show off some of your workflows and how it's improving your workflow and how it's improving your development flow and how it's making your code better, or this or that would be better than like, yes, look at this crazy stuff I can do. But again, it's going to be, I feel like this is almost more of a person problem, right? Like a, how, how do you relate to the people that you are working with to get them to understand the benefit to them without triggering their wall that they have most

• 34:35Scott Tolinski?

  likely built up?

• 34:37Wes Bos?

  I think as you as head of engineering, like the top engineering needs to be constantly sharing how they are using this tool to help them. And because they're probably have, they're blinded by like, oh, look at the one prompt I can do X, Y, and Z. Whereas like, even inside Sentry, there's a Slack channel where just like, everybody is either asking questions about like, how should we do this? Or a lot of like, the top engineering talent are giving their opinions. And a lot of that is critical of it. And then a lot of it is, is really good stuff as well.

• 35:11Scott Tolinski?

  I think that's a great, if you have a company Slack or company communication channel, having some regular means of like sharing things that you've learned there, I think that's a good idea to get in front of people. Cause I do, I love the Sentry one.

• 35:23Wes Bos?

  Yeah, it's great. And even just like a screen recording, like I don't think enough people are showing how they are working with this type of stuff. And just, I know that's type of videos don't do super well, but if you could just internally in your company share, hey, here is my process. I have this bug. I'm working on it. Here is what I'm, here is what I'm doing. And you can show the benefits at the other end.

• 35:49Scott Tolinski?

  Yes. Yes, totally.

• 35:51Wes Bos?

  Question from Andrew says, hi, Santax. How long do you think we will take to move on from using notification channels like Telegram as a primary control for non-coding AI agents? Looking at Hermes desktop and OpenClaw 2.0 looks like web UIs or disguised as desktop apps. Is that where we're headed? Or are dedicated devices like the one Scott built? The lack of purpose-built ways to interact with AI agents is one of the reasons I haven't fully adopted them yet. Other one being the amount of bloat that's being added automatic, automatically, which is why building your own thing is fun, especially when optimizing for smaller local LLMs. I think everybody is trying to figure this out right now is

• 36:34Wes Bos?

  what is the golden goose of UIs to be able to interact with it, right? We saw Codex has merged with ChatGPT and it is simply just one app. Everybody's building these just like super apps. We're starting to see the two big ones right now, GrokBot and Muse from Facebook, which are much more targeted at like, these are not for coding. These are just for your, your daily tasks. And those ones are much more as like a chat application, you know? But then there's, there's not too many knobs that you can turn. That seems to be getting the, the normies built into it. I do agree that the like, the simple Telegram chat back and forth is probably not it because you do need a little bit more control over

• 37:18Wes Bos?

  the UI of these types of things, but not too much where it becomes very overwhelming because, and I think that's why both GrokBot and Muse have been like very heavy on, they don't show thinking trails. They don't show tool calls. They're just, they're just showing like chat bubbles, like you're doing iMessage with a friend. So I think there's probably somewhere in between those things where it's a casual chat, but has a couple knobs where you can adjust some settings. And I think that that will be the winner for a lot of these normies coming on to use this type of stuff.

• 37:57Scott Tolinski?

  Yeah. The totally custom stuff is great because it can do whatever you want. But let me tell you, mine breaks all the gosh darn time. And the missus is not happy about that. I think I get a message once a week, the AI is not working again. So it's something that if you're going custom, yeah, certainly. Like, there's some really great stuff you can do there. But again, you're, you might be running into situations where you have some janky software comparatively. So yeah, there's a lot of other people have built custom ones. If you're using OpenClaw or Hermes or whatever, there's a ton of people who have built custom ones. And honestly, the Hermes desktop app is really good. I don't know if I would put it in front of my wife, but it's good. So yeah, that's where

• 38:40Scott Tolinski?

  I'm at with this. I still have my custom one, but it is kind of annoying. Next one here is from Diabetic Codes. Hey gents, I know you've answered the, should I still learn to code in the world of AI question a million times, but I'm asking it a bit differently. Having been in the industry since 2014 and stepping away for two years during the framework boom, I felt like I've been fighting an uphill battle ever since. Diabetic Codes, you're in luck because no one cares about the framework war anymore. Now in a world of AI, despite having a drive to be a proficient developer, I've worked at two jobs that use AI daily. I'd struggle between taking time on tasks to learn the code base and using AI to get tickets done faster. Additionally, where I used to enjoy digging into learning something

• 39:22Scott Tolinski?

  new in programming every weekend, I now find my drive depleted. I don't want to touch a computer. What's a good way to stay motivated when you aren't motivated? I'm a very disciplined person with fitness, health, and life, but this is one thing I feel I can't find ideas to sink my teeth into and code out with excitement and curiosity. All the best. Yes, this is tough. You know, I actually struggle with this a lot too, where the, it's like, it's like playing a game with Game Genie. I don't know if you had, you didn't have video games, did you Wes? I don't know you had video games wise.

• 39:53Wes Bos?

  Uh, yeah, the like Nintendo, but not much of a video game guy. Uh, my neighbor did. I'm very familiar with the game Genie, yes.

• 40:02Scott Tolinski?

  So Game Genie made the games fun for about five minutes, maybe that. You give Mario some super jump and then you're like, okay, but this makes the game easy, right? This is no longer challenging. And the losing part of the challenging bit of coding where you're digging into like concepts and ideas and you're using your brain in deep ways. Instead of now, you're just like, ooh, agent fix this. A what's wrong agent? Okay, agent fix that, right? You're getting into a different flow where that that lack of real challenge in that sort of way can really negatively impact. But also there's a number of other reasons why AI coding is more exhausting. Whether or not it's just like the the the breaking of flow

• 40:47Scott Tolinski?

  or various things that we heard from people when we did our AI mental health survey. There's so much here. So how do you find the excitement and curiosity? To me, I've found the joy in understanding problems at a different layer. A layer of strategy, a layer of like deep refinement in UI, a layer of of understanding like bigger problems and structures and architectures and things. Because you can tell the AI coding to agent to do this or that, whatever. And there's times when the agent's going to pick a really dumb idea to do something. So like really, for me, it's like I've had to find that joy

• 41:31Scott Tolinski?

  in solving problems in a different way, in challenging myself to solve these problems in a more efficient way, to improve my process, to improve my harness, my skills, these types of things, digging more into that kind of stuff where the code becomes a result of the improvement of my process rather than an improvement of the code by my hands directly.

• 41:55Wes Bos?

  I agree. Well put.

• 41:58Scott Tolinski?

  Thank you.

• 41:59Wes Bos?

  Tim Apple says, not a question, but a suggestion. I think Wes mentioned never having restarted his computer so much as he is now when he's running AI. I've created a shortcut on my Mac to kill all running node processes and close all running programs. This seems to have helped manage the rogue AI agents on my computer.

• 42:19Wes Bos?

  It's not a bad idea. There's a lot of stuff running on my computer that is node based, right? Like my stream deck right now. And if I kill all node processes I have in the past, then the stream deck stops working. And ideally that thing would just restart itself, but it's not not the case all the time. But it this certainly is a problem that will need solving, which is spinning up all these rogue processes. And they're just they're they're headless now and they're running forever. You don't know where they are and they're just churning, using up CPU cycles. Like my battery on my computer, it given it's old. But you could tell that there's stuff running because it doesn't last nearly as long as it used to.

• 43:01Wes Bos?

  I used to be able to get like 14 hours from it and I got like an hour and a half the other day.

• 43:07Scott Tolinski?

  I know. I need better. You know what? I did really like Orca. So Orca is like an app that you can run agents and stuff in kind of like T3 code or any of these things. And the one thing I really liked about Orca is there's a resource manager built into it that shows how many megs and stuff or even occasionally gigs of CPU and RAM your agents are using, all of them. So when you're running a whole bunch of agents, you can see so clearly just how much that stuff is using. And then if you were to run your dev process and stuff through there, I'm sure it would too. I don't I don't like giving my like, I don't know. Yeah, I have some stuff with that.

• 43:49Scott Tolinski?

  But I think just we it's easy to lose sight on all of the the process that gets spun up and restarting your computer definitely more important than ever. But I guess I just in general need a better way to deal with that than right now what I'm doing, which is doing what you're doing and just deleting things. As you notice it getting pressure or restarting, which.

• 44:10Wes Bos?

  Yeah.

• 44:11Scott Tolinski?

  Since I'm running beta software on here, I feel like restarting is probably not a bad idea anyways. All right, last question here we have from B Murph. I work for a company that has invested a lot of time and resources into building our own personal internal harness that sits top Claude code and implements a ton of custom skills and agents. In my opinion, it just stinks. Yeah, it sounds a little iffy. It's a token hog, probably caching, runs slow, runs agents that don't make sense for the given problem and a slew of other issues. I understand the impulse to build a lot of rigor around people using more and more AI, but I feel as though this approach is sort of wrong. Now, for my actual question, is this insane harness what we should be investing time and energy in too?

• 44:54Scott Tolinski?

  Or is sort of the smaller custom agent per problem with humans as the main orchestration the way to go? In my experience, I have much more success giving my AI small problems. And I'm just quite dubious about the YOLO harness. Am I too skeptical, a hater, or in the right ballpark? Hey, I think harness is a tough problem. And harness experts, I think, are there for a reason. There's a lot of harness challenges. You don't want to break caching. You want to have things being routed efficiently if there's model routing involved. You want to make sure that everything is in the place and working the way it needs to be to have an efficient harness. What I prefer instead of a custom harness,

• 45:39Scott Tolinski?

  because you know me, I'm the type of guy that likes to roll my own stuff. I've never considered rolling my own harness because I just don't feel like that's mine. That's not that's not for me. That said, I think custom tooling around a harness certainly makes sense rather than the harness itself. So I think I think there is a layer of your company creating something to around a harness that makes sense. But creating the harness itself doesn't make too much sense to me. So I think that the answer is somewhere in between your idea of like, okay, humans as the main orchestration and smaller agents per problem.

• 46:22Scott Tolinski?

  You know, like I'm going to take care of this full control over this versus the custom harness. I think the answer is somewhere in the middle. I think I think you did a little bit of they went left, you went right. I think it's somewhere in the middle sadly.

• 46:33Wes Bos?

  I also think that all of the stuff we've learned over the last eight months, so much of it doesn't even matter anymore. I was thinking the other day about like contact seven MCP for docs. Do we need that anymore? I don't I don't know that we need that anymore. Because it's just so good at searching the internet and finding the stuff that we want. You know, like I haven't I haven't been running that in a while. And I gave it a library to do 3D rendering for me. And I had tried that about a year ago and it just it couldn't even make like a box before the whole thing fell over. And now it's like it's amazing. It's like rendering full 3D printable stuff. So if you've put a whole bunch of stuff into your own custom harness, you know,

• 47:18Wes Bos?

  you've made your own G stack, maybe it's time to like say which are the pieces that are actually helping us here? And which are the pieces that are just making it slow and token hungry? Because even all a lot of these skills people have installed, they're just adding unnecessary token churn and unnecessary like ways to just spend money. It's it's I even find that with the agent browser, I have to turn it off sometimes because it just wants to take screenshots of everything and it's it's making everything slow.

• 47:47Scott Tolinski?

  Fill my drive with screenshots.

• 47:49Wes Bos?

  Yes.

• 47:50Scott Tolinski?

  Yeah. Awesome. Sick. Let's do sick picks and shameless plugs. If you knew the show, sick picks are just stuff that we've been enjoying lately. Could be anything, Wes. I'm going to be picking a video game, which I like video games. I play on my Steam Deck mostly, so that's typically where where I'm playing my video games these days. I got a new a new game, which is called Driving Rogue, which is a rogue-like driving game. People describe it as like, what if Bellatro was a car game? Which like car, not card, because Bellatro is a card game. Which on the surface, that kind of turns me off a little bit because there's so many like what if Bellatro was blank games lately?

• 48:34Scott Tolinski?

  Because you know, Bellatro is like one of those really cool games. But again, like the the the reward loop of it, I feel like is a little manipulative. So Driving Rogue actually has like skill involved where you have to be good at driving the car and doing the stuff. And then there's rogue-like elements where you have these quick races that are improving your car. It's got a really dope art style. At very first, I was like, yeah, this feels like I don't need rogue-like things in my driving game. But I gotta say, I've been playing it a bit and it was cheap, like nine bucks when I got it. Could be 10 bucks by now, who knows? I gotta say, I've been really enjoying it and having a great time with it.

• 49:18Scott Tolinski?

  So if you that type of thing seems like in your Q zone, check it out. I actually think it's been really, really cool, so.

• 49:23Wes Bos?

  Sick. All right, I have a sick pick and it is the Amazon Basics battery screwdriver. This thing is awesome. I've been wanting to get a battery-powered screwdriver, which is, I know that's what a drill is, right? But I've been wanting to get something just for disassembling for the longest time. And I finally, I watched a video on a YouTube channel called Torque Test Channel. And what he does is he reviews all the like, he buys like the impact driver from every one of these tool brands. And he'll actually like test out how much it costs and the torque on it and if if it breaks, et cetera, et cetera.

• 50:04Wes Bos?

  And he did one on these little electric battery-powered screwdrivers. And I was shocked that the number one by far was not like DeWalt or Milwaukee or Skill or anything like that. It was the Amazon Basics electric screwdriver. And this thing is fantastic because.

• 50:26Scott Tolinski?

  Really?

• 50:26Wes Bos?

  Yeah, it's like good enough to like break free a lot of like nuts. It's it's good enough if you like want to like screw directly into like a piece of wood. I was surprised. I was like, those things probably aren't that good. You know, they probably like cam out and you can't can't do any real work with them. But man, oh, man. And it was like 15 bucks or something crazy like that. Like the the battery in here alone is worth like eight bucks. And I was very surprised at how good this thing was. It comes with a couple of bits and whatnot. And I do want to get some of the like, it takes like the quarter inch bits, which is not for like electronics and like like small stuff.

• 51:10Wes Bos?

  So I do want to either buy another one that's more for like smaller electronics or just buy some adapter bits that will help me. But just for around the house, like one example is we hang stuff up all the time. And I use these amazing like metal screws that go into the drywall. And then you and you screw into that. And I used to have to get a drill, drill a little hole and then by hand crank that in and then screw into it. I can just I can put those straight into the drywall, no drilling or anything. All I need to hang something up now is just this tiny little thing, which I love.

• 51:44Scott Tolinski?

  Cool.

• 51:45Wes Bos?

  Yeah, check it out. Amazon Basics.

• 51:47Scott Tolinski?

  I'm always using my drill and having it set to have too much torque.

• 51:51Wes Bos?

  Oh, yeah.

• 51:52Scott Tolinski?

  For dumb things.

• 51:53Wes Bos?

  You can adjust the torque on here as well. So you know, if you don't want to overdo things, but I just I let it rip. Okay, sick.

• 52:01Scott Tolinski?

  Okay, cool. Well, that's all I got. Anything else, Wes?

• 52:04Wes Bos?

  That's it. Thanks, everyone, for tuning in.

• 52:06Scott Tolinski?

  Peace. Peace.