Very Important Agents  ## Metadata - Author: [[Changelog Master Feed]] - Full Title: Very Important Agents - Category: #podcasts - URL: https://share.snipd.com/episode/1feb63f6-f33f-4758-8436-50c0a84d0410 ## Document Notes ## Highlights - **Working In Public With Vision Pro** - Nick wore Apple Vision Pro at a coffee shop and worked while traveling between cities. - He described connecting via Zoom avatars and using the device during flights with Starlink internet. Transcript: Nick Nisi I before coming home to record this fantastic show with y'all, I was at a coffee shop wearing my vision pro working at a coffee shop. At a coffee shop. You're that guy. I'm that guy and proud to be it. You're the Vim. You're the very important man with this very important Vision Pro on his face. Adam Stacoviak Now, Jared, had I known this information prior to the invitation, I would have considered rescinding. Just rescind that. I mean, we could do it right now. Oh, no. Just like a five-minute episode. All right, y'all. This is the show. That's it. Sorry, guys. Nick dropped the ball. We can't hang out with this guy. Jerod Santo So who gave you this idea and why did you follow it? Nick Nisi I've been doing a lot of travel, which I've been spending a lot of time on airplanes and like trying to get work done. You know, my flight to I've been flying mostly to San Francisco and I've been flying United. And for some reason, the Omaha to San Francisco direct flight, for some reason, has free Starlink, which is amazing. So I can literally like stream TV shows and have faster than that. I could take Zoom calls, but I'm not that guy. Not yet. Jerod Santo You're not. I don't know, dude. We're going to have to reconsider. Adam Stacoviak Might as well. Right. Might as well. Is there Zoom in the Vision Pro? Maybe. Nick Nisi There is. There is. And when you connect with that you connect with your avatar so they just see like a weird 3d representation of me which is amazing is that what we're looking at right now or could we get Jerod Santo No this is really me this is really you did you get the ray-ban glasses yet i have a pair uh i used to work at better not the ones with the glasses were you a glass hole i can't remember if you Were a glass hole or not. I was this close. Oh, okay. Now we know where he draws the line. Or at least you used to draw it. He's an edger, man. He's on the edge. He is. You have all the cool new toys. So when did you get this, Vision Pro? This is news to us. Nick Nisi Yeah, I got it pretty much the day it came out. ([Time 0:05:13](https://share.snipd.com/snip/763bbe78-4213-40b1-ba95-30550a86acc0)) - **Anthropic Buying Bun Signals AI-First Runtimes** - Anthropic acquired Bun but will keep Bun open source and MIT licensed while integrating it into AI tooling. - The acquisition signals JavaScript runtimes moving closer to AI-first developer workflows and tooling. Transcript: Jerod Santo Bun is being bi-anthropic. Yes. Oh, you didn't say it. You're waiting for me. Nick Nisi I said bun. I didn't know if you were going to say Anthropic or Bun, so I was trying to match. I know. That was good. Jerod Santo So Anthropic, the corporation behind such products as Claude and Claude Code. I'm not sure what else they got going for them. Claude for Chrome. Okay. We might talk about that later. That'll come up later for sure. I think they use Bun. I think Claude is like written in Bun style JavaScript stuff. I guess, is there a Bun style? I'm far enough away from Bun at this point. I've interviewed Jared a couple of times, but it's been a few years that I can't remember how different it is from Node. Is it Node compatible? I think so. Okay. So it's like, take your Node code and run it faster with some other stuff they added over. Anyways, I'm flaunting my ignorance here. But I'm pretty sure that Claude code is written in Bunn and then Anthropic because they're flush with all this investor money, I assume. Like raising billions and billions of dollars and they're like we're gonna buy it bun and uh bun said yes which i think is what i would say if i was running bun wouldn't you guys i mean they're Nick Nisi Probably getting a bunch of money so that's a bunch of baloney that sounds uh bun-tastic i don't know what to think oh why not i mean it's exciting right but what what do they need to do to Bun that they couldn't do as like this more open thing or like independent thing what does anthropic need bun to do or to do to bun as you said it if i was a diehard bunner what like is this Good news for me or is this bad news because anthropics preference on whatever they whatever direction they need to take bun into is now where it's going and so does that mean i am left Out because maybe it's not the direction that i'm also hoping my product will go with bun or do i now have like special treatment because now like anthropic or claude will inherently Know everything about bun and know how to perfectly write code for bun applications as just a side effect or was bun ink or whatever, Oven, I can't remember what they call it. Jerod Santo They got good puns over there. Were they on their last couple of months of payroll and they didn't find a way of monetizing? And Anthropic is like, we want this thing to continue. It's open source and MIT licensed. So the project would continue to exist, but not at the pace of progress that a full-time team can put behind it and maybe it's one of these bailout situations where it's like let's acquihire Because jared sumner and his team obviously very talented engineers built something that's quite useful and good and has helped push the industry forward in the javascript server Side world let's use some of this extra cash that we're sitting on because they raised so much money now that being said they're also burning through a lot of it with gpu purchases with Rentals with data center build outs they're doing all kinds of stuff anthropic and we don't know if they're actually making money per token at this point somebody probably knows but I don't know are they still you know losing money every time somebody does an inference i know they're losing it on you nick i know they're subsidizing your account oh i don't know i think I'm subsidizing everyone else's or not i don't know that's that's something that's anthropic is private. So they and their investors know that answer at this point. But if they are flush with cash, they certainly have cash on hand. And they're probably looking for ways to spend it. Seems like a not expensive in there. I mean, expensive for us to buy, but not expensive for anthropic to buy proposition but i definitely get your hesitation like where might it go from here you never know if the ai industry Is a bubble and the bubble bursts you know is anthropic more sustainable than bun is i don't know they seem like they are they certainly seem like they're on stable ground but time will Nick Nisi Tell yeah it's it's exciting i think it's a win for javascript overall i think that bun and dino are serving their purpose of moving node forward and maybe this is just like giving the Torch back to node a little bit on on that because made a lot of great progress. Yeah. I think in the more recent versions, like there's not a feature of bun that I'm like, I can't, I feel like I don't get in node. Right. Maybe, maybe I just don't know. I don't pay attention to bun too much, but like that's, you know, you can compile, you can run job, run typescript like without compiling, which is great. You can compile binaries. You can, ESM still sucks, but everything else seems good. There's like a, what is it, SQLite library built in now? Or coming, at least? That's cool. Adam Stacoviak Gotta have it built in, man. Does it get into a single binary? Is it always a full-flesh full-feature web app? Is there binaries? Can you compile to a binary? Nick Nisi You can. I think that it's like bundling up the JS runtime and all of the node modules into some kind of binary that can run without you having to have node installed. Adam Stacoviak Yeah, in that case, you'd want that SQLite built in because you don't want to have to have reliance on anything else in the system sure you can ship a binary but if you have to rely on sqlite Being installed then that's not cool yeah sure yeah i don't think they're all-in binaries are anywhere near as efficient as like goes, for instance. Jerod Santo Like they're going to be quite large. And I think Dino's is the same way where it's like it's a large binary, even if it's a small program. Now it's a large program. Who cares? But if you want to like, you know, roll out a three megabyte compiled thing, you're not going to get that without some fancy footwork from Bun. And I don't think from Dino either. But the convenience is certainly there yeah which is really most of it at this point i mean we're all very rich when it comes to hard drive space and bandwidth for sure you know earlier Nick Nisi This year i was at um squiggle conf uh that our friends uh josh kohlberg and Dimitri Metropolis run. And there was a speaker there, Oliver Medhurst. I think he was at, they were at Firefox before that, before what they're doing now, which is building a version of JavaScript that can compile to a straight binary without having to have That wrapper. ([Time 0:10:19](https://share.snipd.com/snip/d51dd6c9-6a4f-4389-9151-15a973f00a28)) - **Watch And Trim Model Context** - Monitor context usage (slash context) to avoid bloated conversations and long-running state in Cloud Code. - Trim and compact conversations proactively to stay within effective token windows. Transcript: Nick Nisi And if I just spin up Cloud and run slash context without giving it anything, I'm at 112,000 tokens. Which is insane. You need to trim up that cloud md it's it's that it's the mcp file like mcp tools loading in how many tools are you using not that many it's like yeah nearly you are mcp is hungry it ([Time 0:47:30](https://share.snipd.com/snip/72122fc1-a1c1-46e5-8c75-2a98f2ad886e)) - **Package Reusable Agent Skills** - Publish reusable agent marketplace entries (plugins/skills) so you can share and install agent behaviors. - Structure skills to encapsulate intent, parameters, and subagents for repeatable automation. Transcript: Nick Nisi I do. I have a whole marketplace You can get the Nick Neesey official marketplace I'm just saying. Jerod Santo Now I know why he's here. You're going to show his marketplace. Okay. Let's hear this marketplace skills. Nick Nisi Oh, like they just have this, you can create a Git repo and you put like a marketplace.json in it and it can then, and then you can just say like plugin, add marketplace, uh, whatever Nick Neasy slash Claude plugins, I think is mine. And then you can like install all of these plugins from me that are like agents or commands or whatever and it's just a way for me to like move them out of my dot files and into a a way that you Jerod Santo Can pick and choose which ones that you actually want and play around with it what are some of the skills you may find there yeah do you have any nunchuck skills any go hunting skills i've Nick Nisi Been talking a lot with uh with chris bone skull bone skull hiller i almost said bone skill because he's bone skills that'd be a good rebrand yeah we've been going back and forth on this Because he's he's digging really into this as well is he digging into this oh yeah okay um and the like one that that's really good is uh i have a code simplifier one that like, after we get Something working, this just goes through and it tries to like, in a way, de-slopify the code. So it's going to remove the comments that no human would actually write. It's going to simplify things. It's going to look for things, but it's, it's going to do it, uh, like around the vision that it shouldn't break the exit like the things that work right now it should just clean the code Jerod Santo Up right and so that's like a great thing to run so i just told my claw to do that at times now but does your skill have like details about how you want that to happen is that what the skill is In there yeah instead of saying like hey do a pass and just clean everything up like it actually has details okay and how do you trigger that then i just it's set up as an agent a subagent Nick Nisi So it'll like kick off its own context window and go so i just say ask agent code simplifier to simplify this code or something like that can't you give it a cooler name than that like why Jerod Santo Are you gonna call it agent code simplifier can you like agent X or something, you know? Nick Nisi Well, because it's in my marketplace now, it's like at Nick Neasy colon essentials slash. I don't know. Like, yeah, it's a whole thing now. But you could call it whatever you want. You could take it and copy it and call it whatever you want. That's a cool one. I also made one that is a, um, uh, it's like a consultant, uh, because like, I also want to experiment with these other models and figure out what's, what's going on. And so I made a consultant that can do three different things, depending on how you ask it. The first one is you can be like, ah, I don't know, you know, if Claude is like spinning its wheels on something or get stuck, I can be like, ah, that doesn't sound right. ([Time 0:48:12](https://share.snipd.com/snip/8ed8ad2e-5ae9-4f78-8f58-084f714863c4)) - **Team Of One, Many Cloud Instances** - Nick runs many Cloud Code instances simultaneously and manages them with tmux. - He built a tmux dashboard showing project statuses and Claude instance states to avoid constant manual checks. Transcript: Jerod Santo Yeah oh yeah love it all day now how much of this is work all of it it's all work he's contractually obligated to say that all of it i'm currently a team of one uh so I make up for that by having Nick Nisi A up to, I think I've had up to 14 cloud code instances running simultaneously. And I use Tmux to manage it all. I built like a little Tmux floating window dashboard that I can pop up and it shows me like, you know, this project is waiting on a response for me. This project's done. And like, it shows me the Claude status of each one. Jerod Santo Do you ever feel like you're babysitting? Yes. Nick Nisi Constantly. Jerod Santo Yes. Constantly. I bet. ([Time 0:53:21](https://share.snipd.com/snip/dd4787ed-2282-42c1-b57c-a04959037968)) - **AI Browser Experiment Fell Short** - Nick tested multiple AI browsers (Atlas, Perplexity, Comet) and found unreliable results when asking it to locate a referrer. - The experiment left him feeling these AI browser UIs are currently ad-driven and brittle. Transcript: Nick Nisi I opened up ChatGPT Atlas. I have all of them on my machine right now. I've got Atlas. I've got Perplexities Comet. I've got Dia, that piece of crap. Adam Stacoviak That piece of crap. Nick Nisi I've got Chrome with Gemini built in and claude for chrome and the whole the whole thing but uh i opened up chad gpt atlas uh and it just so happened that the last time i used that i had opened Up my fathom analytics page uh for my website so when i opened it up it brought me right back to that, which is like, was interesting. Cause I hadn't looked at that in a while. And I noticed that like one of the referrers was Claude.ai and it was specifically to a blog post I have about like, you should try Claude skills. And I was like, Oh cool. Did they link to me? And I was like, trying to figure out how to get there and see like if i could find a page where i'm linked or something and i couldn't so then i just popped open the the sidebar in atlas and Said hey i see that they're a referrer here find the page where this came from and it popped in and it's like yeah we'll help you with shopping and it had like like it was just so stupid um And it of course did not find anything so i don't i think it might have been like come up you know come up in a clawed answer or something like that but it totally failed and it jumped right Into their their failing as like you know an actual ai provider they're just a a way to sell you ads is the way that I felt as soon as it said chat.jpg shopping. ([Time 1:01:59](https://share.snipd.com/snip/8108b117-0157-4212-a6ed-c5a37f05358a)) - **Chats Could Replace Visiting Websites** - AI chat UIs can embed live app views (maps, Zillow) so users never visit the original website. - That dynamic rendering in chat suggests browsers may become secondary to agentic interfaces. Transcript: Nick Nisi And it will, and under this budget and it like loads like a Zillow page in there, a Zillow map. And you can move around the map right in your chat window. You didn't go to Zillow.com. It's like loading that and helping you find it right in the chat there. And then you can have a contextual conversation around that. I feel like that's where things could be going. And like the browsers just not figuring out like where they belong in this is indicative of that being the future. ([Time 1:04:26](https://share.snipd.com/snip/2028b9e0-49c1-4ecb-a07f-e4091a4219ed)) - **More Output, Less Manual Coding** - Developers will produce more output but write fewer original lines as AI handles routine code. - The role shifts from typing to reviewing, designing, and validating AI-generated work. Transcript: Nick Nisi Yeah um i think that like the code the code that we're writing is if you're still writing a lot of code, you're going to write less next year. I think that that's a given at this point. Uh, I think there's more, you're going to produce more, but you're going to write less. You're going to be involved less in the day to day minutia of the code. ([Time 1:27:22](https://share.snipd.com/snip/ddc853be-e01c-46a7-a0fe-00331321919f)) - **Use Output Styles To Share Coding Work** - Use output styles (system-prompt variants) to make AI handle repetitive code while you focus on business logic. - Configure the model to stub mundane parts and pause for core decisions so you stay involved and accountable. Transcript: Nick Nisi Yeah um i think that like the code the code that we're writing is if you're still writing a lot of code, you're going to write less next year. I think that that's a given at this point. Uh, I think there's more, you're going to produce more, but you're going to write less. You're going to be involved less in the day to day minutia of the code. Uh, there's ways around that to be as productive, I think. Like one thing, I don't know. You asked for a prediction. I guess that's my prediction. Can I give you a recommendation of something to try? Adam Stacoviak You can do what you want. Nick Nisi All right. So I'm talking about Claude code because that's the one that I use the most. But you could probably do this with the other ones. But Claude has this thing called output styles that they deprecated. And then they got pushback on that. And so they undeprecated it. And so it's still here, at least for right now. But it's a way to change the system prompt a little bit so that you can tell it exactly how you want it to respond to you. And it ships with two examples. One of them is explanatory where it'll like give you like a little insight bubbles about why it's making a decision that it made or whatever. And the other one is learning. And that learning one is specifically for, you know, I want to go learn the go programming language. And so when you have that enabled, it's going to optimize for you to fill out like some of the more like, Oh, it'd be good for you to learn how to do loops in go. So I'm going to stub out this method and you're going to fill it in. And it'll like give you prompts and tell you how to, how to do it in like code comments. And then you can, it'll pause there and have you do it, but you can ask questions before moving on. You can tweak that. You can make it in as many output styles as you want, but you can tweak that and you can tweak it to be more of like a business logic-y one. I forgot what I exactly call mine. It's like, yeah, I can't remember what I call it. Basically fill out the minutia of the code, like stub out this method, add the getters and setters and all that stuff that is boring, you know, just get the work done work, have it do that. But then like the core business logic stuff, pause and let me do that. And we'll talk about it and walk through it together. And you can even do most of it maybe, but like that way you're more involved in what the code is actually doing. You're stopping to take part and participate because otherwise the code is just scrolling by you. And then, you know, already like in the past, you know, six months go by and I have no idea what code I wrote six months ago that I have to go debug. It's like that week to week now with all of this code. So if you're not keeping yourself involved in it, then you don't really have a solid footing mentally on what code is passing by you and that your name is being attached to because it's Still your name on the commit. ([Time 1:27:22](https://share.snipd.com/snip/a9f344b1-c8c3-497c-852b-eea9b9b6d902)) - **Validate AI Bug Findings With Repro Apps** - When an issue arrives, run an automated check to identify likely causes but always validate manually. - Keep vanilla example apps to reproduce bugs and confirm AI-suggested fixes before committing changes. Transcript: Adam Stacoviak What is your flow? How do you truly pay attention to the code? Give me the very, very detailed visual, exactly what's happening. Paint it as best you can. Nick Nisi Experimenting so i play around with i've been playing around with um stevie eggy's thing um is it beads i think yes yeah looking at that i've been looking at um github github's thing specify And i just do my i rolled my own as well and the problem with rolling your own is you have to maintain it and as things models get better are you still using the best practices like i just don't Want to chase that constantly so that's why i'm like experimenting with other ones because then they can go chase it and i can just reap the benefits but i have been playing around with That also opus 4.5 i don't know if this is an opus thing specifically but like their planning mode is different now and it's a lot better it asks you questions yeah like it asks you questions In like an interface it saves that off to a dot dot like in your home directory dot cloud slash plans and then it gives it a random name for the plan but it's there as a markdown file and i noticed That and it was asking me to do all of these things it was saying like run slash plan it kept telling me that yesterday run slash plan slash plan does not exist so it's like giving me a preview, I think of what's to come, but I play around with all of those. So a feature comes in, if it's like a GitHub issue, I have it, like, I have a command that I run to have it read the issue. And it's usually like, let's look at like, is this a legitimate thing? And so if it is like, it'll run through that and be like, oh yeah, I see that it's on this line of code where I think the problem is. And that's like how I do a lot of reactive stuff is like that. But I don't just trust it and say like, oh yeah, that's it. I have like example apps and I, that are just like vanilla, you know, they have work OS auth kit set up in them or whatever. And I just have them, I'm like, okay, let's create a test page in that app. ([Time 1:36:13](https://share.snipd.com/snip/100e1857-ff35-4f68-b186-fa218304fe8b))