OpenClaw Makes AI Agents and CPUs Get Real

just tell cloud code your problem. That's how I started. I was like, look, I have a lot of newsletter emails. I read a lot of news sources. So, what do I do with all of this? Give me a workflow that sorts all this out. So, it said like, yeah, I'll give you a slash news skill. All you have to do is run /news and it'll create the digest for you. I'm like, yeah, awesome. Do that.

Hello listeners and welcome to another Semi-Doped podcast. I'm Austin Lines of Chipstrat and with me is Vic Shaker from Vick's newsletter. Today we're going to talk about Agents.

So, the framing for this is I feel like 2026 is actually going to be the year of Agents. I think even as far back as 2024, there started to be some marketing speak where lots of companies said Agent AI is all the rage and it's the thing and everyone should get on the bandwagon. And I think a lot of that was very aspirational and visionary, but it feels like now, as we'll get into with things like Cloud Code, Cloudbot, and so on, I'm actually seeing lots of people tinkering with Agents. So, more than just marketing on a PowerPoint, but I'm actually seeing people play with this stuff.

So I thought Vic and I we could kick off the conversation chatting about Agents and then we will get into what are some of the implications for CPUs for GPUs and so on from an infrastructure perspective. So, okay, enough of an intro. Vic, let's start with you. Agents, what's top of mind for you?

Yeah, so the Agent thing has become quite a revolution on X. Like, everybody has declared that it's over and now the age of Agents is upon us. And also, I'm not sure what we should call this tool that is now this great Agent that everybody's talking about because it started off as Claudebot, but then Claude decided it sounds too similar to their product and then they changed it to Moldbot and then now it's again changed to OpenClaw. I think this is what the final product is going to be called. So maybe we should refer to it as OpenClaw so that this podcast makes some sense even a few months down the line.

Yes, totally. Okay, we'll call it OpenClaw today. We'll see what if it changes.

So this I saw this really funny YouTube video. He's a software programmer. I've seen his YouTube video for years and he calls himself the primogen. So, you know, if anybody's watching this, you should go and pause this for a moment and see his Claude Code. Not his Claude Code, his OpenClaw video that he has. It's really funny.

But essentially, it's just like all these Agents are doing random things like it went and lost money on the crypto market and then posted on LinkedIn, I lost $1,600 in the crypto market and here's what I learned.

So, it's ridiculous. So, this kind of stuff is really hilarious. The last week has been awesome. Just simply awesome in the world of technology and LLMs and AI.

That's good. It's so LinkedIn clickbay type sounding, you know, like I did a thing and now I'm an expert. Let me tell you about it. That it's almost mockery. It's hilarious.

And it was all posted on this like Reddit equivalent for Agents called Moldbook. So these Agents go and they post on Moldbook like humans do like they're writing a LinkedIn post. It's just a mess. I'm like why are we why are we spending all this infrastructure in GPUs and burning down a few rainforests so Agents can get on social media like as if it weren't bad enough for humans?

Totally. Yes. the experimental stage we'll say that yes indeed indeed.

So let's start with Cloud Code which when I think of like agentic AI coding I think of Cloud Code. Tell me are you using Cloud Code? How are you using it?

Yeah so the thing is with Cloud Code for the longest time I never gave it a whirl because I don't do programming for a living and so I was like what I mean I guess it's a coding tool whatever I don't need it because I'm not coding anything. And I know programmers have been using it for a long time and doing it's been doing a lot of good stuff for them. So I was like I don't know I never use it.

But then recently I saw videos of stuff and people talking about how Cloud Code is so much more than just coding and then people use it for other things. And the creator of Cloud Code himself said this on an interview. So I was like okay maybe now is the time to do this and I'm not scared away by like a Linux terminal or whatever. I'm okay with terminal. So I was like, fine, let's fire this up. It's pretty easy.

So I all I did was, you know, installed Cloud Code based on the instructions on the website. It works right off the bat. So it's on my Mac, you know, I installed it on my Mac terminal and I use it on that. So it's pretty neat actually. So I have the $20 subscription and I've been using it fairly extensively even based on that. I at one point I got locked out of Claude Opus. And I've been using Opus the whole time. I could have used Haiko or Sonnet. I didn't. I was like I just want the best model. Let's see how much it goes.

So almost for a week's worth of work for my use case, I think the $20 was enough, right? So I got some use of it. Like I've been starting to create some skills and like running it and helping it do some of my research. It gives me information quickly. So yeah, it's pretty cool.

Okay, say more about how you're using it and which let let me pause really quick and say that it's you point out like a a funny like branding problem. It's like called Claude Code. So you're just like ah I'm not coding so I don't need this thing and they're it's kind of like no no no wait it can do more which I know that anthropic has kind of tried to get past that with Claude Co-work I think it's called putting into the guey where you know normal people live let's say you know and giving it a lot of the same tools have you and and so I guess another question too so tell me more about how you use Cloud Code but have you thought about also using that in Cloud Co-work or do you just stick with Cloud Code because Do you like the terminal command line feel versus a guey?

I'm okay with the command line feel actually. I like it. So I'm so I actually use that. The cowork aspect of it I have not yet gotten into because a lot of the files that I work with at least for the newsletter and stuff is on Google workspace and Google drive. So I don't really need to access local files as much.

But I can see if like cowork you know if you give it some control over your computer you know it could reorganize files on your computer this and that kind of stuff. But that being said I really don't like giving machines control over my computer.

Because you know this thing is watching me. I feel like I'm being watched and I have my password manager which has like a master code and like everything in my life is past that. Totally. I'm like, I I'm not going to type anything while my Agents are watching me.

I mean, think about it. If you hired interns to offload work, you're not going to let them use your computer, right? You're going to give them their own computer.

Cloud Code isn't running, you know, it's it's within the the terminal window that I'm using it, but it doesn't have general access to my computer. If I kill the terminal session, it's it's out, right? like now no more.

So for you it's almost like a sandbox thing too. It's like hey it just lives in the terminal. It can't see everything. It does its thing there and then I can kill it when I'm done.

So the way I'm using it is actually I have hooked it up with a Google Drive MCP server. So I run that and on a local Docker container on my Mac. So you can get download Docker for desktop. I think it's there for Windows, Mac or whatever it is. And then you can get the Google MCP container from the Docker registry and you can just run it.

And what that does is it creates a link between the Google Drive on you know Google server and your local Cloud Code session and MCP if people haven't heard about is Model Context Protocol. So it's kind of like an API for LLMs that you can plug into. So with this thing, I can actually look at my emails. I can look at my files on Google Drive and I can look at my calendar. So I can ask Cloud Code, tell me what my week looks like. It'll look at my calendar and give me an, you know, something like that. I I could go look at it, you know, whatever. But I could also never leave Terminal and be like, hey Claude, you know, can you add this appointment at this time? And it I don't have to switch windows at all, right? So it just goes and adds a

Okay, interesting. So yeah, I was going to ask like, oh, why are you connecting to Google Drive when you could just put files locally, but it's not just Google Drive, but you're accessing your calendar and your Gmail. So that that's pretty awesome.

So that's a very good point. Actually one thing that I ran out of tokens doing is trying to put the files that Cloud Code creates like let's say I ask it to research something for me and create a report or something just so that I can read it and understand all the news of the week. I have a skill like that. But then what it does is it spends all this token authenticating with the MCP and I have to copy paste this thing into a browser and I have to click okay there. Then I was like, Oh man, I'm just like dumb. Like Austin figured it out in like 1 second. I should just save it to my local Google Drive folder and then that will sync.

So that's then I asked Cloud Code like, Hey Cloud Code, don't sync it to the online version. Here is the folder locally. Just put it there and it'll sync on its own. So it does it like it it then takes care of the changes, right?

Nice. Yes. Good. Good. So you you mentioned you've got like a skill for reading the news or summarizing.

Yeah. So what I told it is like I have all these Substack publications that come through. By the way, when I say this, people are going to do this to our our publications. So it's probably not a good thing, but it helps, you know, it helps. The reason it helps is that I just want to get a sense of what the article covers without having to open all of the Substack publication emails I get and then have a quick summary of what is in there and then I can go and read the whole thing if I find that it's relevant or what because I store it on a Google doc or a Google file drive all these news summaries it now becomes searchable because Substack search is so terrible.

So if I just go and ask Claude, hey, look in my news folder and tell me the substack that I remember seeing sometime in the last couple of weeks that spoke about this particular thing and it will surface it for me and there is the you know it short summaries there too and now I can go read the original thing. So it's kind of a note takingaking thing.

So what I asked Cloud Code to do was I have all the Substack emails coming into my Gmail folder labeled like newsletter and then I asked it to go read all the emails tagged newsletter and then it automatically decided to tag them as newsletter processed you know and it tag tags it another tag yeah it says like I'm going to tag it process because if I do this next time I will not you know resurface this because it's already summarized for you, right? So, it's very clever that way.

And it said like, look, I'm going to put this newsletter process tag as hidden in your Gmail so that you don't always see it because not relevant really. So, it put it it created a new tag label and hid it. It's very beautiful. It makes really good choices.

Yeah, it's really thinking deeply for you.

Yeah. Okay. Interesting. So, it's reading your email to get access to all your Substack, which is good because I was thinking like, oh, do you have it authenticated to Substack? How does this work? Technically, it can only read the entire email. So, for all of the Substack newsletters that are too long and get truncated, it's obviously going to only read what's in the email, but then it's it's even cleaning up after itself for you and and tagging it as processed and hiding it. That's pretty amazing.

Yeah, it's pretty great. And now the thing that you mentioned about authentication is interesting. This is kind of unsafe what I'm doing here. But I have certain new subscriptions that I pay for and those things I I just gave Cloud Code like the login and password and be like log in on the website and then get me all the information on the paid articles because you know digit times is one example right because if you go to digit times and you don't have a pay subscription a lot of the articles are just completely payw wall like there's not even a single sentence over there so this thing goes logs in and then reads

Wow. And then summarizes it for me. Amazing. And it and it and it works. It works great.

And I told it like what I'm interested in too because I don't want all kinds of news articles. So I said, Hey, in the following areas like like I like photonics, I like memory. I like interconnects. Whatever that I think like semiconductor like capex like semicap stuff ASML that kind of news foundry stuff. So it made it, you know, you can tell it like here are the topics I'm interested in and then it goes and creates its own files and keeps it all organized and deletes lines and adds lines whatever and keeps it all like great.

So it only surfaces the kind of information I want, not everything because then it becomes useless, right?

And so is it writing all these summaries locally? It's like a markdown file or something.

Yes, exactly. It writes it in a markdown file on my local Google Drive folder which is which I read with VS Code. Maybe I should find a better markdown here but I just read it in VS Code. But you know if you go to Google Drive it doesn't maybe I don't know there's like a plugin I try to install that reads markdown but it's just cumbersome because it's not a Google doc file and you can open a markdown file as a Google doc file but it's just too much work. I mean it's already on my computer. But I just like open it and I'm okay.

Totally. Yeah. I use some app locally called Bear, I think, uh to to view markdown. But hey, maybe maybe Agents will drive markdown to get better support in like Google and whatnot. Become a first class citizen if you will.

Yeah. I mean it's fine. And then I told it like one, you know, all these like small optimizations I did later because I was like it tries to go through the it creates the entire digest and then it realizes it can't connect to the Google MCP server and then it's like I don't know like it's just hanging out and did write any files and so I told it next time make sure you can authenticate first before you start doing anything. If nothing happens let me know. I'll fix the authentication then you can go read my substack emails you know so I save tokens otherwise I like I just waste time with token like thinking through all this stuff and then it can't write the file you know totally totally so it's both a genius but also can't think for itself and you have to tell it like hey you're optimizing for tokens here

Yeah sometimes it's really smart sometimes it's like ah why didn't you do the obvious thing but for somebody who's thinking of like doing this kind of stuff I would not even recommend like going through any guide just tell Cloud Code your problem that's how I started I was like look I have a lot of newsletter emails I read a lot of news sources so what do I do with all of this give me a workflow that sorts all this out so it said like yeah I'll give you a slashnew skill all you have to do is run slashnews and it'll create a digest for you I'm like yeah awesome do that

nice amazing that's pretty awesome I it's interesting because to you to generally today or yesterday it used to be you know anytime you want to start on something you like look for the tutorial and you hope someone wrote a good tutorial otherwise you're fumbling around and now you can just ask Claude and let it essentially guide you yeah having been trained on all sorts of previous tutorials it's a probably a good teacher at how to walk through and do this stuff

yeah I also created a couple of other skills I said like I have a slash topic skill so what it does is it looks through all my news articles and looks at all the stuff I've written and if there's any possible extensions so it gives me a few suggestions like hey you should look at this look at this has happened in the news you kind of looked at your newsletter because it can go read at least the article top you know whatever is before the payw wall on my you know newsletter itself on substack right that part is open so it goes and reads substack and it reads the news folder on my Google drive and tries to form some links and stuff like that playing with that it's not really useful in a sense because it's very it gives me very obtuse ideas. I don't know. So maybe I have to work on this a little bit. It's not as good as I would like it to be.

Gotcha. And how do you decide how much time to kind of sink into this and if it's worth it?

Yeah, that's a good point. I have spent time on these kinds of things and it goes nowhere. like in the initial GPT days I tried creating various like what was it called? It is like a skill but it's like your own subgp so to speak. Yeah, I tried making those but then you know I couldn't keep them up because each one had a system level instruction and then it didn't evolve with my style or anything like that.

So here I like this because it's like an all-in-one thing and I improve it as I go. like I don't spend too much time on it. Like if I'm doing something and it's coming up that I'm doing the same thing a few times and I ask Lord Code to go make that a skill or something and then it doesn't maybe do a great job of it which is fine because at least it got the job done for now and later on I'll kind of add skills or tweak the skill and all of it happens like I'm telling like telling you you know like oh you know I don't like this do that you know change this change that so it goes and fixes it.

So because you've been tinkering from the early days of making like custom GPTs or whatever they called it to now with Cloud Code and it it sounds like now you feel like you can talk to it more in natural language and it's mostly helpful and sometimes it burns tokens but generally speaking you can actually start to like automate some workflows. Um do you feel like we're starting it back to the year of Agents theme? I mean, do you feel like we're starting to get closer to a place where people who want either are less techsavvy or don't want to invest so much time tinkering, you think we're getting to a place where they can actually start to create Agents and tool around and experiment and find value?

Yeah. So, yes and no, right? Because right now you have these things like OpenClaw which takes Cloud Code to the next level right and it is not as simple as people think if you're not used to like dealing with a little command line stuff here and there and the only reason I say that and let's go into OpenClaw for a little bit now because first of all anybody listening to this don't try OpenClaw on your laptop because literally when you see the website, it tells you all the stuff it can do and it's terrible. Like it's like handing you know unlocking your laptop with a password or your Touch ID and handing it to somebody else. Would you ever do that in real life? You would. So that's that's actually what OpenClaw does, you know. So don't unlock your computer and give it to an LLM. It's crazy to do that. Because it has everything like on your laptop.

So the way people have been doing this is they've been buying these Mac minis because they're kind of powerful machines. They have fairly large number of resources and maybe you could even run a local LLM on it. And we'll come to why that is a little bit later. But then if you don't want to spend like I don't know like hundreds of dollars on a Mac Mini. I don't know what kind of Mac minis these people are specking up on it. Could get expensive.

But the simplest way to begin is to go and get a virtual private server, a VPS. You can get those like for like $5ish and you'll get a like a little Ubuntu terminal and there you can go and you know claw code or not Cloud Code OpenClaw and install that and get going there. Right? So that's that way you on a virtual machine somewhere in the you know cloud at least you don't have all the stuff that you have on your laptop and it's a little better off.

So yeah it sounds like you're saying basically again security comes to mind and sort of sandboxing. You don't want to run it on your local computer. So why not either run it on a local server people are buying Mac minis or just run it on a cloud server as like a safe environment.

So I would say that the Mac Mini has another advantage which is that um you are really not exposed to the outside world if you choose to not be out you know if you choose it to not be so because that's your entire entirely in your control. And what I mean by that is when you put up a a virtual machine on a cloud server and you run it and those machines unless secured properly are accessible to anyone because you can SSH into these virtual machines. You can hack virtual machines because they're on a cloud. They're exposed to the internet. So unless you have proper hardening practices even those are unsafe, right? So you have to be very careful like how people can get into your OpenClaw instance and that's something you have to look into.

I found one suggestion I can give anybody who wants to try this. It's little bit expensive and then I'll tell you how I do it. I' I've been messing around with it. I'm not really using it that much but I'll tell you the way people might want to look at it if they're serious. And that is Digital Ocean has a a preconfigured droplet you can get which is called Moltbot. Stay still using the middle name here. But yeah essentially you can pay them I think $24 a month and you can provision this machine that they tell you is safe to use. It's hardened for like internet access. So that's a good way to do it.

But the costs don't end there, right? because you will hook up this thing to some form of LLM and you can pay anthropic to do this and you will pay API costs and so that's the thing here if you are paying $24 a month already for this droplet on digital ocean that is like secure and then you're paying API cost like another 50 75 bucks a month it comes up to $100 a month which is of course like for the capabilities that you have out of this thing is actually very cheap. But you know maybe not everybody needs this like people like me are like still happy with Cloud Code unless you need some kind of massive automation that Agents have to do this stuff you know you can get by with claw code for now like I'd say so

so clawed bot or OpenClaw or whatever OpenClaw the the benefit is as a pair compos or compared to Cloud Code is it just that you can run like lots of Agents simult simultaneously or what's kind of like the main value prop?

Yeah, so you can run a lot of Agents and you I have seen people propose that like various Agents spin up do a particular task spin down then two hours later and different task spins up does something spins down. This kind of stuff is better than keeping an always on Agent which sucking up tokens. So you wake them up and make them to work and put them back to bed. you know that kind of thing is better.

So you can have sub Agents doing a lot of stuff. So I have seen some people's like ideas of doing this where they have like a manager for their business. This is their manager Agent and under that manager the manager Agent takes care of all the input from the sub Agents like you know a social media expert, a content analyst or an SEO expert. All these people report report to the general manager who runs as an Agent and yeah this kind of stuff is like fancy actually. So the question is how much do you personally need that stuff but if you're running a business that makes sense.

I can totally see why people would want to do this. Right. Sure. Sure. Interesting. Fascinating. But but for you today do you feel like cloudbot's worth it?

I think it's worth playing around with. And the way I set it OpenClaw, the way I set it up is I have my own like home server. I run a little server with with with all my I run my little cloud storage here for like family photos and such because I have a lot of them. So I spun up a virtual machine there and it doesn't have access to the outside world at all because you have to VPN to get into the home server network. So it's all like not exposed to the outside world in the sense nobody can get into it. But it does have access internally to the internet. So I I have been playing around with this. I want to see what it's capable of because what I feel right now is that I am only limited by imagination here because honestly for what I do like clot code is just something I'm even starting to use now. it. I can see how it's helpful. I don't have the sense of imagination to see what I can do with a tool as powerful as like setting up a virtual company of Agents that runs my business. I I have to think about this. I think everybody needs to have that like vision. What do I really automate and you know what kind of work do Agents need to do? And then you can kind of go off and like figure out how to get it done. That's the easy part. Yeah. figuring out what you wanted to do is the hard part.

I mean, I think that's going to be a a new muscle for a lot of people to develop, which is thinking like a manager. You know, how what are the outcomes I'm trying to accomplish and how can I create tasks that can be delegated so that I can accomplish my outcome.

Exactly. You got to have like some amount of self-awareness of what it is that you are doing. Yeah. and how to agentify it. Totally. Totally.

So, I like the idea of having like a local AI server that you can run a lot of this stuff on that's you can control how exposed to the outside world it is. Now, at the end of the day though, everything is an API call to claude and it's consuming all these tokens and you use Opus. it's only 20 bucks a month, but on the other hand, it's and it's the best model, but you you are running out of tokens and presumably if you're like a gentifying your business and you're running tons of tokens, maybe it's going to get expensive. Um do you ever see a world in which the some of the AI is actually like the lighterw weight tasks are actually offloaded and run locally on an AI server or do you think this is always going to go to the cloud?

No, you can do it locally like like the Mac Mini approach or like you mentioned earlier the DJX Spark which is a better inference system you can run on your local premises. Yeah those kinds of things are great like what's what's a few thousand to somebody who's running I don't know a million dollar business right? A few thousand for a DJX Spark I don't know how how is that how much it costs? Well, the the um the Station is like $4,000. I think the Spark the Station has like a GB10 in it. The Spark, which is supposed to have like a GB300, I think those they haven't said, but it's probably be like 50,000 bucks or more.

Okay. So, yeah. So, you can get the small one or the big one depending on your needs, I suppose. But yeah, in either case, if your business is running like several million, which is still a small businessish, you know, it's not a billion dollar business. Even a small business, a medium-sized business, can get one of these $50,000 machines and kind of run a bunch of Agents because how much would an employee how much would a bunch of employees cost? Way more than $50,000, right? Totally. If you were going to hire someone, somebody like this should look at really buying one of these computers, putting it on premises, and running your inferencing on it if you don't want to pay the API costs. So, it all depends on your use case. If you have enough load and enough token usage, go get the API or this one depending on

Yeah, totally. And I'm I'm very convinced like I know that for AI labs it's always about like unlocking the next level of intelligence and increasing the IQ and stuff and and therefore that would be the argument like no no you should always want the frontier IQ and then you should keep using our API. Um, on the other hand, if you were to just run all this stuff on your own machine using some like open source models, which are always getting better, too. They might just be like two or three years behind, like I have a feeling for a lot of this stuff, like a lot of your use cases, it's probably good enough like reading the news and summarizing it, you know, surfacing ideas in front of you, things like that. Like, probably good enough to just do the open source stuff. And then I think you could totally control your cost because you're like, Oh, I invested five grand up front or 50 grand upfront and now I can experiment and go crazy and generate all the tokens possible and that's it. You know, I'm just paying whatever some small monthly fees are for various cloud services.

Yeah, there's another in between option actually between running it locally versus a frontier model. I for my own cloud bot OpenClaw experiments I went and got an open router account which allows you to get a whole lot of other models like deepseek and you know all of the quen stuff you know there like a whole lot of other models that are like not ultra expensive and then I got like $5 in credit just to try this out and I also limited how much this can use. So, I put like upper limits in case I charge like a hundred bucks. I'm like, you know, I have $100 in credit, but I'm like, use only $10 a week, okay? Like, don't use all the hundred in the first week. And so, the first thing I did was I set API limits before I go ex get all excited in agentic world and like blow all my credits in like one day. So, the first thing I did was set limits. I think something that people should keep in mind as well. But yeah, you can get other models at various price points per token. You don't have to go frontier or local, you know, you have a whole range here.

So, you're saying the advantage for open router is you could sort of cost optimize and and try to play around with like maybe telling your Agents, you know, hey, use these models for these tasks only use this frontier model for this type of difficult task.

Yeah. You can even give it a preferred hierarchy of models to go through and say like you know for some things like this use this model and that model. So within OpenClaw you can give it a hierarchy of models to go through. So you can really finally control how you spend in inferencing. Interesting.

This will be a story we'll have to watch because if if 2026 is the year of the Agents, meaning people are starting to actually do useful stuff, pretty quickly behind will come lots of cost optimization of like, cool, I this is awesome. I'm spending a ton of money, so maybe I need to play around with like, yeah, stack ranking models by performance and price and thinking about local versus uh, you know, cloud compute and so on.

Before we get to talking about the implications from an infrastructure perspective more deeply, I I wanted to ask really quick, one of the things you said earlier was um, you know, oh, I've got, you know, I'm using Cloud Code and I've got it reading my substacks by going to my Gmail and summarizing them for me and then now I'm just referencing this summarized file. You also talked about being able to have it go into websites like Digi