915: $200mo Background Agents, CLI Tooling and “Max Mode”  ## Metadata - Author: [[Syntax - Tasty Web Development Treats]] - Full Title: 915: $200mo Background Agents, CLI Tooling and “Max Mode” - Category: #podcasts - URL: https://share.snipd.com/episode/4f780c6d-c058-46bf-8a6e-c12c79fc401e ## Document Notes ## Highlights - **Power of Background AI Agents** - Background AI agents run tasks concurrently and autonomously in the cloud. - They enable set-it-and-forget-it workflows for complex or multiple tasks simultaneously. Transcript: Wes Bos Background agents, long-running AI tasks, some of these coding tools that are significantly more expensive and significantly more autonomous, being able to run it on your own, some That are fine-tune models. Codex, Claude Code, the new Cursor background agents, they just released a new $200 a month plan. And I feel like we're on the cusp of another wave of these tools and another wave of what they can do and how we're actually using them, which is hilarious. Because if you go back to the episode of us at the end of last year, one of my predictions was, I feel like the AI is going to get like, like, I feel like we're not going to see any major improvements This year, you know, we're just going to get significantly or a little bit better here and there. But man, this stuff is really exciting. And I've dove into all the different tools, we're going to talk about how they work and how you might want to use them. And are they really that much better? And is it really worth spending 10 times the amount that you're spending on like a co-pilot subscription or something like that? Scott Tolinski Yeah, one thing I think we did get right was that the AIs themselves might not go through some extreme evolution, but the way in which we use them will evolve. And I think already through June here, we're seeing so many different new, totally different new approaches to interaction with them. And I was actually just got back from JS Nation and React Summit in Amsterdam, where I was emceeing both days. And I had a lot of conversations with people about the interfaces in which we interface with AI. And that was like such a huge topic of conversation, more so than like, will the models get better? Will this get better? That whatever, it's more or less like, how are we interfacing with these things? And I found there to be a massive amount of thought-provoking ideas and discussion there. Wes Bos Because at the very start, I was looking at all these CLI tools, like OpenAI Codex, Cloud Code. Those are probably the two big ones where instead of using an editor, like a cursor or VS Code, you're simply just typing into the terminal. And I was like, is that really a better interface, just typing into a terminal? You don't get the full UI and all the diffs and all of that stuff. And granted, there are connections you can use in it. But the amount of people that said, you're thinking about wrong, Wes, you're thinking about it like that you're just sitting there and looking at what it does. And, and these are much more of a set it and forget it and let this thing go off and run. So that's what we're going to talk about today is this sort of a new wave of everything. Scott Tolinski Yeah, yeah, it's interesting. Yeah, I'm curious to hear some of your thoughts on some of these things, because I did use Cloud Code for a bit. And yeah, I don't know if I got it. So I'm interested in getting it a little bit better. Wes Bos To each other. The first one being this idea of background agents, and the second one being CLI tools, which can be used as just part of your existing tooling. So first one, background agents, they're becoming very popular. The idea with them is that you give them a task, right? We have agents already. You give it a task, you type in the box, and then it goes ahead and does the work that you want. And then you look at it and say, accept or reject, or you type back and forth to it. And sometimes you'll go back and forth with several prompts. The idea with background agents is it can do a lot more. These things are going to have significantly ([Time 0:00:03](https://share.snipd.com/snip/95a56444-fd65-4c4b-92a9-3bf1bdd68e73)) - **AI Migrates Podcast Files** - Wes had an AI agent migrate over 900 podcast episode files updating RSS feed URLs. - It ran 20 minutes doing multiple precise steps with only minor clarifications needed. Transcript: Wes Bos Yeah. So an example might be on one thing that I gave it. I said, add related episode embeddings feature, meaning that like I want it to go off and create an entire feature for the syntax website where it takes every single episode, all the Transcript, it does embeddings, and then it will save those to the database. It will create an API endpoint for those embeddings. It will be able to use those embeddings to find related episodes to each other, as well as create a whole bunch of Svelte components that we could pop like in the syntax website, say like, If you like this episode, here are three more that are similar to this. And then we could also have like a 2D graph that would sort of group them together. And then theoretically, we could use that for tagging the episodes as well. So that's a fairly significant feature, which may take an AI agent like i don't know 30 minutes an hour to go ahead off and and do so what you can do is you can give it a we'll talk about this Like requirements document or tasks document and it can go off and run that and then you can go to another one which is like we simply just have like lots of like little one-off issues on Github that are i don't know maybe half hour, 45 minutes worth of work. And we are beta testing the Slack integration right now where you can simply just set it at a GitHub issue and it will go off and fix those things for you. So background agents, they're a little bit different than just regular agents in that they will run for a larger amount of time. And simply by using background agents, and I've also been testing the like max mode in Cursor as well, it seems to simply just like check its work and just kind of keep recursively going Back over itself until absolutely everything has been done. And it's pretty impressive. Like another example was I needed to migrate. We switched the RSS feed for the Syntax podcast. We moved from Libsyn to Megaphone. And part of that was we have 912 markdown files that have MP3 URLs in them. And I needed to take the old RSS feed, line up the GUID, the guaranteed unique identifiers. I needed to line that up with the new RSS feed. And then I needed to find the related episode markdown file, parse the front matter of that, switch out the MP3 URL, and then push it to GitHub, right? And that's like seven or eight steps. And the thing ran for like 20 minutes, and it did an absolute perfect job of getting them all in for me. Maybe not a perfect job. I had to step in two or three times and clarify what it was. But if I had been a little bit more upfront and a little more clear with how it worked, and there was a lot of like kind of messy data, some of the URLs didn't follow the same patterns as others. And it did like a really nice job where I could simply set it off. I went and went and worked on something else. And then I came back to it and it had done 20, 25 minutes worth of work in the background. ([Time 0:04:31](https://share.snipd.com/snip/db67319f-49c6-4036-bfe2-acbcc6f929e8)) - **Integrate CLI AI Tooling** - Use CLI AI tools as modular primitives integrated with your existing workflows. - Assign AI agents tasks via API links to GitHub, Jira, or Linear for automation. Transcript: Wes Bos And then the second category I have here is CLI tooling. So Cloud Code is probably the biggest one in this space right now. And then there's OpenAI Codex CLI. And then DAX is working on another one, it looks like, called OpenCode, which is kind of cool. You'll be able to just run it and I'm assuming put your own models behind it. So these are just straight up CLI tools, which you can integrate it into your editor. But the benefit to simply being a CLI tool is there's no dependency on your IDE or any sort of integration. So theoretically, you can think of it as just being able to NPM install something. Of course, you can use it via the CLI, but it also is simply just an API, which could be part of something else. So you would imagine you have an integration with like GitHub issues, you have integration with your Jira or your what's the one everybody loves so much because it's local first linear, Linear, yeah, linear integration. And the idea I think with this, these type of integrations, you simply just assign it to an agent. And then that will then go off do its work. And you can build stuff on top of these as well. You can make your own software and being able to layer in adding MCP servers to this type ([Time 0:12:46](https://share.snipd.com/snip/14f5bfea-0c46-4c85-96d6-c752dacd7935)) - **CLI AI Pay-As-You-Go Costs Add Up** - Wes tested a pay-as-you-go AI tool and found costs accrued rapidly with many requests. - He prefers fixed-price plans to avoid stress over usage expenses. Transcript: Wes Bos Yeah. Yeah. I went into it. And let me tell you, I went as pay as you go, which is not the way to go. It's so stressful. Every single keystroke I made, I was so concerned about it costing too much because it adds up pretty quickly. So, for example, I went into a very simple demo that had maybe seven or eight files, I think it was a simple react application. And I just asked it, what does this repo do? Nine cents, nine cents out the door, you know, like, that's not obviously not a lot of money. But if every single time that you hit enter, you're paying 10 cents out the door. And then another prompt I gave was add a new example called iPhone.html with some CSS to replicate an iPhone layout. That went ahead and spent 8 cents, which, again, not a ton. But imagine it going off and it went back and forth three or four times. But then imagine going on like an actual real mega code base and simply it understanding how does this code base work? Which files do I need to include? What does your database model look like? How do I run all of these things? I just ran a linter. Now let me take the linter and the TypeScript output and let me send that back. You're going back and forth. I think like some of my examples with background agents were, I don't know, back and forth 100, 150 times, which can really add up. So pay as you go stresses me out. I'd much rather just pay the single amount and then be able to just blast it and and i was surprised at how many people were willing to pay to i still i haven't shelled out the 200 bucks but We're part of the century cursor um and david kramer was just like let her rip like you know like this stuff is this is this is game changing stuff i personally So he's like, let me know if You need more more credit or whatever. Just like go nuts with ([Time 0:14:22](https://share.snipd.com/snip/668f5eba-ce3f-4eb3-9733-2977fbff699c)) - **Recursive Self-Checking in AI Agents** - Background AI agents continuously self-check and recursively improve their outputs. - Recursive revisiting enhances quality before completion in complex multi-step tasks. Transcript: Wes Bos Single amount and then be able to just blast it and and i was surprised at how many people were willing to pay to i still i haven't shelled out the 200 bucks but we're part of the century cursor Um and david kramer was just like let her rip like you know like this stuff is this is this is game changing stuff i personally So he's like, let me know if you need more more credit or whatever. Just like go nuts with it. So I am still impressed by how many people are willing to pay this much money, which I think I understand. What do you think? 200 bucks. Is that too much? Scott Tolinski It depends on who's footing the bill. Yeah, it is and it really if it's coming out of my post if it's coming out of my paycheck pocket then i would say maybe too much but if it's coming out of like like if i still was a business owner And it was a business expense and as part of operating my business then that's a different conversation for me. Wes Bos If you're paying, like, I guess you have to see, like, the actual quality output of this to see, is it worth it? Scott Tolinski You know, and you have to see some kind of return. Wes Bos I think we're, we are, we are past it being a cool demo, but not quite at the, like, holy crap, this thing is, is amazing, you know like it still does bad things and still you have to babysit It but i'm curious if somebody were like let's say you're paying three or four devs would you pay an extra thousand bucks a month probably i don't know if you it's hard to say yeah and that's Scott Tolinski That's hard's hard. That's a hard conversation anyways. Yeah. Because sometimes you'd rather employ people for a number of reasons beyond just the effectiveness of it. Wes Bos I'm not talking about replacing them, but I'm like, would you give them this $200 a month tool? And I guess like corporate probably would, because the amount that these companies spend just for you to be able to zoom each other it's it's quite a bit yeah no kidding yeah it is that Scott Tolinski Is always mystifying to me when you get into corporate accounts on anything just how expensive pay by head is the is the real money maker on these things yeah let's talk about approaches Wes Bos So what i've learned with doing these background agents is you can't simply just YOLO, type in the box, make this thing. You need a very good planning document for it to follow because it does its best when it's able to have very clear inputs, very clear defined tasks to do, very clear outcomes and success. And also, I found that a lot of them will start to log their work as they are doing Markdown. So what was happening with the migration of the RSS feed is that it would loop over and it would log, OK, out of the 912 episodes, 115 didn't work. Let me now evaluate those and see if there's anything that I can, oh, I see. These ones had a different URL structure. Let me update the regex so that it would match a little bit better. And it put all of that in a markdown document and then reran it and then put it the updated. Okay, now out of those 80 more were successful. Now let's look at the last 20 that we need to do. So it's putting it all into a markdown document. So what I found is for it to be able to sort of tear off a task and work on that specific task and then work on the next thing, I think you get much ([Time 0:16:01](https://share.snipd.com/snip/9f8f803c-029a-4c0b-a8e3-6a46248fd2e2)) - **High Cost Reflects High Compute Use** - $200/month AI agent plans reflect the intense compute resources required. - Businesses may find value if the AI accelerates development compared to human costs. Transcript: Scott Tolinski It depends on who's footing the bill. Yeah, it is and it really if it's coming out of my post if it's coming out of my paycheck pocket then i would say maybe too much but if it's coming out of like like if i still was a business owner And it was a business expense and as part of operating my business then that's a different conversation for me. Wes Bos If you're paying, like, I guess you have to see, like, the actual quality output of this to see, is it worth it? Scott Tolinski You know, and you have to see some kind of return. Wes Bos I think we're, we are, we are past it being a cool demo, but not quite at the, like, holy crap, this thing is, is amazing, you know like it still does bad things and still you have to babysit It but i'm curious if somebody were like let's say you're paying three or four devs would you pay an extra thousand bucks a month probably i don't know if you it's hard to say yeah and that's Scott Tolinski That's hard's hard. That's a hard conversation anyways. Yeah. Because sometimes you'd rather employ people for a number of reasons beyond just the effectiveness of it. Wes Bos I'm not talking about replacing them, but I'm like, would you give them this $200 a month tool? And I guess like corporate probably would, because the amount that these companies spend just for you to be able to zoom each other it's it's ([Time 0:16:41](https://share.snipd.com/snip/560166bb-1083-483f-94c9-25f594cf0cf0)) - **Use Clear PRD Documents** - Provide AI agents with clear, structured product requirement documents (PRDs). - Clear inputs and defined tasks improve AI outcomes significantly. Transcript: Wes Bos Type in the box, make this thing. You need a very good planning document for it to follow because it does its best when it's able to have very clear inputs, very clear defined tasks to do, very clear outcomes and success. And also, I found that a lot of them will start to log their work as they are doing Markdown. So what was happening with the migration of the RSS feed is that it would loop over and it would log, OK, out of the 912 episodes, 115 didn't work. Let me now evaluate those and see if there's anything that I can, oh, I see. These ones had a different URL structure. Let me update the regex so that it would match a little bit better. And it put all of that in a markdown document and then reran it and then put it the updated. Okay, now out of those 80 more were successful. Now let's look at the last 20 that we need to do. So it's putting it all into a markdown document. So what I found is for it to be able to sort of tear off a task and work on that specific task and then work on the next thing, I think you get much better outcomes than rather just saying, implement This feature. And hilariously enough, a really good way to make these documents called PRD. I didn't even know what this was. Product requirement document, which is you can create a really nice structured PRD document by just kind of spitballing all of your ideas of what needs to happen. Throw that into chat GPT or something like that, and then get back a really nice structured document, and then give it a once more, once over, clean it up, make sure it does everything You want. Because the cleaner your input is... Scott Tolinski Yeah, go ahead. No, to be clear to the ([Time 0:18:27](https://share.snipd.com/snip/d3eecb2b-866a-4040-ae2a-77520e3d13fb)) - **Find Quality Used Projectors** - Search for quality used projectors for affordable, reliable high-lumen options. - Avoid cheap Amazon projectors which often overstate brightness and perform poorly. Transcript: Scott Tolinski Wes. I haven't spent a ton of time with background agents just yet. I'm still like in the baby agent. You've been more of a foreground guy right now. Been more of a foreground guy. Yeah. But no, I've gotten into cloud code a little bit, but that's about it. I'm going to try some of these background tasks in Cursor today and get busy with some of that stuff. Do you want to head into sick pics and shameless plugs? Stuff you want to sick pic? Wes Bos I got sick pic for you. One sec. Me too. Scott Tolinski As he dips below his desk. Wes Bos He dips beneath his desk. It's not like I was going to come up in a costume. I was just unplugging something. Scott Tolinski Oh, okay. Oh, a little projector. I love projectors. Wes Bos This is my sick pick for a 12-year projector that's actually really good. 12-year projector. We have, at our cottage, we have a projector because we don't have room for a TV, but we have like a huge wall that we can project on. And I've had a BenQ, whatever, projector I bought right before the pandemic. And unfortunately, it is starting to die. And it's not the bulb. It's the chip inside. And it's not worth fixing. So I was like, all right, I need to find a new one. I know you have like a really nice laser one you can play outside and whatever and I was like I kind of want one of those but I also don't want to spend any money yeah so I was like I gotta find Like another cheapie on marketplace and whatnot and I've been looking for a while and I found this one which is a Dell m2 10x and it's i don't know it's like 10 years old or whatever and i Found it and these were really big in like corporate world so there's lots of them on the market right now and it's bright as hell 2000 ansi lumens which is bright enough to have like in Like daytime be able to project not outside inside yeah inside during daytime and it's got hdmi inputs it has audio in and out and it's just like i got for 50 bucks and 50 bucks 50 bucks and Like man if you want like a good i initially bought a like 200 one on Amazon. And I plugged it in. I was like, this thing sucks. And it said it was like 8,000 lumens or something like that. And then I very quickly realized that projectors ([Time 0:25:08](https://share.snipd.com/snip/bf10dcf1-bbf4-4632-847d-16482d18b64f)) - **Smart Ring Beats Smartwatch Bulk** - Scott uses a Ringcon smart ring for sleep and health tracking with long battery life. - The ring offers less bulk and subscription-free tracking compared to smartwatches. Transcript: Scott Tolinski Got a a smart ring which oh you know what people are always talking about wearing the apple watch to sleep and stuff like that i just cannot do that that thing is too big and bulky yeah man I like i sometimes like lay on my wrist and stuff i just can't do it yeah sometimes I wake up with no feeling in my arms. Wes Bos Yeah. Scott Tolinski Well, the big thing about the smart rings to me has always been the subscription fees and stuff like that. I cannot justify another subscription. So Courtney did a bunch of research because she wanted to track her sleep and she got a Ringcon Gen 2. And I had never heard of Ringcon before she did that research on it. She got this one specifically because it does tracking of sleep apnea and stuff. And I've had it now for I got it right before I left for Amsterdam. And it's just been lovely. It's been lovely to track sleep. It does a really good job. The battery lasts for like 11 days. Oh, man. And then it charges and everything like that. And it, you know, it's calculating all your stress and all those things. And for me, I've been very stressed out. So like, being able to look at that and be like, Oh, I need to like, really get it together right now has been very nice. But overall, yeah, I, you know, I do say for like, like just general like vitals and whatever tracking, I found this thing to be much nicer than having the watch. Wes Bos It's honestly amazing that the tech, ([Time 0:31:10](https://share.snipd.com/snip/06cec435-bfb2-488c-8e71-0bd88558baeb))