Episode Categories:
Resources:
Join the discussion on LinkedIn: Got something to say? Let us know on LinkedIn and network with other AI leaders
Upcoming Episodes: Check out the upcoming Everyday AI Livestream lineup
Connect with Jordan Wilson: LinkedIn Profile
Start Here Series in our Inner Circle Community: Join for free access
Anthropic Claude’s “Computer Use” Feature: A Pinpoint Review for Business Leaders
Anthropic’s recent viral release—its “computer use” feature—earned 75 million views online, with attention-grabbing demonstrations suggesting the dawn of agentic, AI-driven knowledge work. But does this functionality genuinely move the productivity needle for business owners and decision makers, or is it just a technical spectacle? This article unpacks the exact value, limitations, and tactical implications based on a deep hands-on review, bypassing hype and getting to what actually matters for enterprise workflows.
“Computer Use” for AI Productivity: What It Actually Does
The “computer use” feature allows Claude’s models (currently via paid plans on macOS) to control a user’s desktop environment by mimicking basic user interactions. This includes opening applications, moving the mouse, typing, filling out forms, and navigating across the desktop. Unlike prior integrations limited by API or web-based access, this update grants AI the ability to interact with any application, even those lacking direct cloud connectors or command-line interfaces.
When enabled, Claude can perform tasks that traditionally required manual input—such as opening proprietary software, running scripts, or copying data—by seeing and controlling the screen and keyboard. However, the access must be deliberately permitted per app, and significant security considerations emerge since the feature can “see” anything open on the shared screen.
Integration and Setup: Key Business IT Considerations
Deployment of the “computer use” feature requires downloading the Claude desktop app for Mac, activating the function in a discrete settings menu, and granting specific accessibility and screen recording permissions at the OS level. There is granular control over which applications the AI is allowed to operate, but organizations should note two key factors:
The process is not silent or invisible; full security and compliance reviews are strongly recommended before enabling on company-owned devices.
Every new app request typically triggers a permission prompt, with recurring re-authorization required for each new session or run, currently limiting potential for set-and-forget automations.
These procedural and security factors make “computer use” more suitable for sandboxing or individual experimentation at this stage, rather than for full enterprise rollout.
Dispatch and Remote Control: Remote Work and Business Continuity Use Cases
The companion “dispatch” feature, released shortly prior, extends these capabilities to remote or mobile scenarios. By pairing a mobile device with the desktop, a user can prompt Claude to “go fetch” files or execute simple tasks on their work computer from anywhere. A typical practical example: sending an important file stored only on a desktop machine while physically away from the office, a scenario that previously depended on human intervention or at-home technical support.
For stakeholders with distributed teams or hybrid work policies, this hints at streamlined offsite workflows, lessening certain “human duct tape” scenarios—that is, manual intermediation for simple, repetitive digital tasks.
Performance and Workflow Limitations: Details Every Team Must Know
Despite the headline-grabbing viral demo clips, real-world functionality is bounded by several concrete constraints:
Speed: The feature is notably sluggish; every action is executed via iterative screenshot-analysis-act cycles with substantial latency. For high-frequency or time-sensitive operations, this remains a bottleneck.
Concurrency: While Claude is exercising control, it “hijacks” the mouse and screen. The machine cannot be productively used for other tasks at the same time, unlike agentic AI browsers that isolate their activities.
Instruction Following: In more complex or multi-step workflows, instruction adherence breaks down. Claude sometimes defaults to using built-in cloud connectors or Chrome extensions rather than actual computer controls—even when explicitly instructed otherwise.
Browser Interaction: Browser automation (e.g., filling in web forms, navigating SaaS platforms) is inconsistent and prone to errors, falling short of established browser automation tools often already in use by businesses.
Authorization Fatigue: Each new application or new session prompts another explicit permission grant, curbing the utility for automated or scheduled overnight workflows.
Security and Privacy: Enterprise-Level Risks and Guardrails
By enabling “computer use,” the AI is effectively granted privileged access, able to see screenshots of the entire desktop and interact with sensitive data. There is the risk of accidental exposure if confidential materials are visible while the feature is running. Anthropic’s own prompts urge users to start with trivial or easily reversible tasks, highlighting the experimental nature and the inherent risk should anything go wrong.
For companies considering pilots or wider experimental rollouts, robust protocols are needed:
Restrict use to non-sensitive environments
Rigorously control which apps and files are accessible
Ensure clear user policies and audit trails
Tactical Value for Businesses: When Claude’s “Computer Use” Is Actually Useful
For knowledge work dominated by proprietary, legacy, or siloed applications lacking automation hooks, this feature finally offers a bridge for AI-driven workflows—albeit in a slow and sometimes unreliable format. In environments where investments in direct integrations or command-line workflows are not achievable, “computer use” may deliver incremental productivity.
Yet for the substantial share of enterprise workflows already integrated with modern APIs or cloud connectors, using “computer use” is explicitly overkill and inefficient—slower, more resource-intensive (measured in AI “token” usage), and less reliable than existing automations.
Strategic Outlook: AI Duct Tape for Human Duct Tape Tasks
While currently impractical for automated, high-stakes, or mission-critical processes, the architecture of “computer use” points toward a future where AI agents can finally move beyond sandboxed SaaS environments and tackle the full heterogeneity of real-world business software. Until the broader SaaS ecosystem universally supports agentic protocols, AI-driven desktop navigation is expected to be the next “duct tape” layer—bridging legacy gaps that still demand manual effort.
Businesses seeking real productivity improvements should treat “computer use” as a practical research preview: suitable for pilot experimentation, limited-scope process automation, and exploratory development. As upstream providers like Microsoft, Google, and OpenAI race toward similar features, keeping informed and evaluating hands-on will help capture early value (and mitigate risk).
Key Takeaways for Business IT Leadership
The “computer use” feature is immediately beneficial only for workflows not already automated via cloud, API, or CLI.
Security, privacy, and manual re-authorization must be built into any pilot deployment.
Performance and reliability are not yet sufficient for enterprise-scale unattended process automation.
As desktop AI control becomes standardized, expect rapid iteration and broader integration across enterprise platforms—requiring teams to stay vigilant, sandbox new tools, and prepare for a phased transition from “human duct tape” to “AI duct tape.”
By understanding the exact mechanics and limitations of Anthropic Claude’s “computer use,” decision makers will better align expectations and investments with the real state of agentic AI for everyday business operations.
Topics Covered in This Episode:
- Anthropic Claude Computer Use Feature Overview
- Viral Launch of Claude Computer Use
- Claude Computer Use Setup Instructions
- Desktop App Integration and Security Concerns
- Functionality: Opening Apps, Mouse, Forms
- Dispatch and Mobile-Desktop Workflow Explained
- Computer Use vs. Native App Integrations
- Limitations: Speed, Screen Hijacking, Permissions
- Live Demo: Claude Computer Use in Action
- Instruction Following and Agentic Browsing Issues
- Agentic Workforce Potential and Future Outlook
- Comparison with Microsoft, OpenAI, and Google
Episode Transcript
Jordan Wilson [00:00:17]:
In their new computer use tool, Anthropic just released one of the most viral AI launches ever and what could be a step change, not just for the AI industry, but for the future of knowledge work. And if I'm being honest, one of the biggest pieces of long term job security for us human knowledge workers right now is the fact that we're the ones opening the programs on the computer, clicking around, saving and uploading, you know, the stuff that I call human duct tape on this show. Because for the high value important tasks, many of us are already relying on AI to do the heavy lifting, and that kind of relegates us to just the human duct tape that opens, closes, copies, pastes, uploads, and downloads on a computer. But now Anthropic's new computer use tool that they just rolled out can kind of do those very things. Like an actual human behind a computer, computer use in theory can open programs, browse the web, access your local files, and use a mouse and keyboard to navigate around just like we do. So as we put AI to work on Wednesdays in our weekly segment here on everyday AI, we're gonna tackle the new computer use feature from Anthropic and answer the question if it's a just a fun party trick or if it's a real agentic workforce. Alright. Here's the big picture.
Jordan Wilson [00:01:44]:
Claude's new computer used tool went absolutely mega viral on Twitter. I'm talking about 75,000,000 views, which is a lot. Yeah. If you're not on the Twitter machine, what people call x, I call Twitter. Yeah. This is one of the most viral AI product launches ever at least by attention and eyeballs on the product. And Anthropic showcased the ability to literally control your Mac desktop via your phone. Yeah.
Jordan Wilson [00:02:09]:
You can do this all with your phone right now with a combination of the dispatch dispatch feature, which they released earlier in March, and the brand new computer use functionality. So computer use can be used on your mobile phone via the dispatch app, the dispatch feature in the Claude app, or it can be used on the desktop Mac app via CoWork or Claude code for paid pro and max users. But it does obviously require full use of your computer. So we'll talk about that later, some of the downsides. So this might not be, kind of an AI agent that sits down with you and works, alongside with you like you might want it to be. So stick with me for the next twenty ish minutes. And on today's show, you're gonna learn why the most viral AI feature right now is maybe much less useful than you might want it to be. You're gonna learn what Claude's computer can actually do versus what it looks like it can do.
Jordan Wilson [00:03:10]:
And you're gonna know whether this is a genuine agentic workforce tool or just a research preview gimmick. Alright. Let's get into it. If you're new here, welcome to Everyday AI. My name is Jordan Wilson, and, well, this is an unscripted, unedited livestream podcast and free daily newsletter helping everyday business leaders like you and me not just keep up with what's happening in the world of AI, but how we can make sense of it to grow our companies and our careers. So welcome on the journey. If that's what you're trying to do, it starts here. But the cheat code is at our website, youreverydayai.com.
Jordan Wilson [00:03:46]:
Go side up there. We're gonna recap the most important things from today's episode, as well as give you all the other AI news you need to know to be the smartest person in AI at your company. So let's get in and talk a little bit now about the new computer use feature. So here's, well, what it can actually do on your computer. It can open apps. It can move your mouse. It types. It can fill out forms.
Jordan Wilson [00:04:12]:
It can navigate your entire desktop if you give it access to. Right? We're gonna tackle, some of the security, considerations here in a little bit. But right now, this works both inside Claude Code and Claude Cowork on macOS. So this is a desktop app that you download and in doing so after you, you know, go through all the permissions, it can literally control your entire computer, not just the Mac app, not just the Claude app. Right? It is available right now for paid plans only, so you do have to be on a paid plan. And if let me just cut to the chase. This is only truly useful, I think, for apps that don't have a dedicated CLI connection. Right? Command line interface.
Jordan Wilson [00:04:58]:
If they don't have, a useful API, if it doesn't have, like, an MCP or if it doesn't have a direct cloud integration. Right? Which that sounds like a lot of ifs, but it's actually not because I would venture to say that the overwhelming majority of, kind of apps or where we work. Right. Because the reality is as knowledge workers, all most of us really do is we're inside apps. Right? You're inside Salesforce or Slack or Outlook or, you you you know, looking on different websites. Right? You're you're you're in Excel. You're in Microsoft Word. All, for the most part, we do.
Jordan Wilson [00:05:39]:
We spend all of our day in apps. Right? And many of those apps do have a direct integration with Claude. And if they do, using this computer use is overkill. Right? At that point, it is just a party trick because it is slow. Right? It's slow but effective. But if you're just thinking, oh, well, you know, I can just kind of put together everything that I I use and just do it all via the computer use, that would be extremely token inefficient way to tackle this. Because, yes, it is number one, slow. And number two, it does eat up tokens like I eat pancakes at the original pancake house.
Jordan Wilson [00:06:14]:
Like, it goes through them so so quickly. So if you do want to set this up, here's how you do it. Alright. And, podcast audience, you know, sometimes our our new Wednesday, series that we do putting AI to work on Wednesdays, Sometimes they're a little more visual. Right? But if you are following along just on the podcast, I wanna do my best to describe what's going on on the screen. I am gonna do some live demos here in a little bit. But right now, if you wanna know, well, how this works, how do you get it to set up? So number one, you have to download the Claude app if you don't have it already. It's only available right now, for Mac, so not Windows yet, and you do have to be on a paid subscription.
Jordan Wilson [00:06:55]:
Alright. So if you check those boxes, then what you're gonna do inside your settings. But in it's the settings for the desktop app, which is different than the settings if you go to clawed.ai. So you're gonna wanna open your desktop app, look in the settings on the left hand side. It's kinda hidden down there, so you're gonna wanna look. There's specific desktop app settings. You're gonna click on general, and then there's going to be a new computer use toggle that you want that you're gonna wanna click on. There's other things you can do from there.
Jordan Wilson [00:07:27]:
You can go to denied apps. Right? So if there's certain apps that you absolutely don't want Claude to access, you can add those apps there. And as you go along, and we'll see, I'm gonna do some of these things live. We'll see how that goes. Because, yeah, it's it's slow, and I've done a couple, you know, practice demos on this, but it could go off the rails. That's the other thing. Right? You do need to be careful. And so this is not something I recommend that you kind of, quietly or sneakily, install on a company computer.
Jordan Wilson [00:08:00]:
I wouldn't do that. Right? If you are a decision maker at a company, you do have to say, is this something that we want to do? I would say it's at least worth sandboxing, but you do have to keep all of these security implications in mind. Because when you do give a large language model complete control over your entire computer, obviously, a lot of things can go very, very wrong. Right? I do these things as a small business owner myself. I understand the risks. Right? I, kind of keep tight guardrails around private, or sensitive information when I do these types of things, but keep that in mind. You can, add denied apps. And then once you do that, you do also have to go into your Mac settings, and grant access for accessibility settings in screen recording.
Jordan Wilson [00:08:45]:
So that essentially allows Claude to perform different actions on your computer and to see things, right, that happen on your computer. Then dispatch, I think, is where this actually becomes useful. And we'll get into that a little bit more as we do the demos. But, if you don't know dispatch, dispatch was released a couple of weeks ago. So Entropic has been on a complete tear and they've actually released so many things. It's kind of confusing to keep up. Alright. So dispatch was released about two weeks ago, and this is just a way to, communicate with Claude Cowork from the mobile app because there's not a dedicated Claude Cowork app on the phone, but there is a dispatch, kind of feature that allows you to talk to Claude CoWork and talk to Claude Co.
Jordan Wilson [00:09:37]:
So So you might saying, okay. I'm confused. Why would I wanna do that? Well, as an example, right now, on your mobile phone, there's no way to access something on your desktop. So you go through with dispatch. You go through a one time pairing process. But then essentially, right, I think of some instances where this would have been very helpful. Right. It's actually funny the the, kinda little demo or use case, that Anthropic led with was, you know, hey, you know, using dispatch.
Jordan Wilson [00:10:04]:
I'm running late. There's this file on my computer. Can you send it over to me? Right? That's happened so many times. Luckily, my wife bailed me out. I was doing a a big keynote in training, in Chicago last year, and I'm like, oh my gosh. I, you know, I don't have this file on my laptop. It's saved on my desktop at home. It's not in the cloud anywhere.
Jordan Wilson [00:10:23]:
Right? So, my wife, who is awesome, bailed me out via it was the pre dispatch dispatch. But now, right, as long as you have the app, running on your home, quote, unquote, home computer, your desktop computer, you can access any file just via dispatch, but that's not computer use. Right? So the way that it does this is when you set it up originally, well, it can just run things through the command line. So it's it's kind of taking a backdoor approach, again, within the guardrails that you set up originally. So little confusing, but dispatch is just a way that your Claude app on your iPhone and your Claude app on your Mac can kind of talk to each other and access files that you've already set up, via folder access. But now the combination of dispatch and computer use is actually really cool. So there are these now new settings or features inside dispatch. So same thing on the desktop app, you're gonna wanna click on dispatch.
Jordan Wilson [00:11:27]:
And then there's a couple new options, on the right hand side, and one of them is enabling computer use. There's also the option to, keep your computer awake and allow all browser actions. And I don't know why the allow all browser actions is actually a toggle, because no matter how many times I tried to get this to work, you still have to always, click allow, like, a trillion times, which takes away from its utility. And I hope that anthropic, fixes that portion, because it's extremely buggy in that regards. So that's a, short, not very short, primer on the setup. So and then when you do that, anthropic does kind of warn you. Right? So it verifies, and I'm just gonna read this prompt that, Anthropic gives you. It says turn on computer use.
Jordan Wilson [00:12:15]:
Claude will take screenshots of your screen and control your mouse and keyboard. You'll approve each app, but not confirm each step Claude performs. Keep in mind, some actions can't be undone. Apps you approve could open other apps that you haven't approved. Websites and docs could contain malicious instructions that misdirect Claude, and it says close anything sensitive. Claude can see your screen. And then as a reminder, at the bottom, it says, this is a research preview. Start with tasks where mistakes are easy to fix.
Jordan Wilson [00:12:48]:
And I think this is a good, kind of call out by Anthropic. Just saying straight up even you can approve certain apps, but those apps might open other apps. And, maybe show an example of this. We'll see. I don't want this to turn into, you know, another thirty five minute show when I say stick around with me for twenty minutes. But even in instances where I tell computer use to use a certain app, sometimes it's like, nah, don't want to. I'm gonna use a different app. Alright.
Jordan Wilson [00:13:16]:
So we'll get into that a little bit more. But here's, how Claude's computer use actually works. Well, it takes a screenshot of your screen, decides what to do next, acts and repeats the loop. So, yeah, it's slow. Alright. And you can choose, which model to use, and, you know, maybe there's a trade off there. I've been using Opus, a little bit more because when it comes to, agentic and computer use benchmarks, it's the best, but it's also very, very slow. Right? I'm starting to test some, different workflows in Sonnet and in Haiku as well, but it's pretty slow because it's essentially just using computer vision, and it's screenshotting everything on your computer, and then it's, you know, kind of, directionally slowly moving the mouse.
Jordan Wilson [00:14:02]:
So you do need to watch for the orange outline on your screen. So this is a newer feature, and this shows you that Claude is using computer use. Because sometimes you can even ask it to use computer use and, well, it might not. Right? So even as an example, Claude says, if, if you have different connectors, already connected like Gmail or Slack, and you say go to gmail.com and log in and do a, b, and c via computer use, it might do that sometimes, or it might actually just default to using your connected, integrations inside Claude. So if you do kind of look for that orange outline on the screen, that confirms to you that, yes, Claude, in this case, is using computer use, and it's not just running in the terminal or using right. One thing I found, when I tell it to use, you know, an agentic browser. Right? Like, I'm trying to use, you know, computer used to control, perplexities comment or to control Chanteevities Atlas. When I do those things, it will open them, but it'll actually just do the task in Chrome because it has a, Chrome extension, the cloud Chrome extension in there.
Jordan Wilson [00:15:15]:
So even when I tell it, hey. Go open, you you know, ChatGPT Atlas and do this task. Well, it's like it'll open ChatGPT Atlas, but then it's gonna decide to do the task elsewhere, and not use computer use really at all. So you do have to kind of monitor it. So, yeah, those connectors run first. Chrome sometimes run second even when you don't want it to, and full desktop control really only kicks in as a last resort sometimes. So I do hope in the future that there's a quite literal, a toggle that you can put on that forces, you know, the quad app to use computer use by default versus it going through connectors or Chrome first. And then, like I said already, dispatch is what makes this worth considering.
Jordan Wilson [00:16:02]:
Because one of the downsides, which I'll talk about that more in a minute, is you can't really do too much on your computer when you're using this. So I think right now, one of the best, use cases for this is when you've stuck away from your computer. Right? So, being able to use the dispatch the dispatch app to essentially do any work that you would normally want to do in front of a computer. I'm one of those for me. I don't know if it's because my fat fingers or because I'm used to having all these big. Right? I usually have, you know, these two twenty five inch or 24 inch screens that I'm always working on, in multiple desktops on those screens. I I I I stink. I absolutely stink at doing anything on the phone.
Jordan Wilson [00:16:41]:
So, I do like using the dispatch feature because then, essentially, it's like, okay. Well, I'm using my entire computer now wherever I am. Right? And being able to dictate that with my voice is really nice. But here's the limitations that no one's really talking about. Yeah. This thing got, you know, 75,000,000 views online, but no one's talking about the things that don't work very well. The biggest thing is it just screen hijacks. Right? So a lot of times when I'm working with different agentic browsers, they kind of leave you alone.
Jordan Wilson [00:17:16]:
Right? Let's just use as an example, perplexity's comment or chat gbt atlas. Right? If they are doing agentic browsing, well, they're gonna take up that portion. This is different because it uses your mouse. You can't really do anything. So I can have even just a lot of different agents or agentic browsers kind of going on their own desktop on my computer or their own screen or their own monitor. You can't really have that with cloud computer use because its screen hijacks, it mouse hijacks. So if you think that, oh, this thing's gonna work alongside me on on one side, and I'm gonna do all my stuff, you can't really do anything while computer use is running because it needs your, it needs your mouse. It needs your screen.
Jordan Wilson [00:18:01]:
So you can't really do anything else. The other thing, it is slow. It is finicky, especially when you're using multiple monitors. Right? We'll see when I do my live demo if that works. And the instruction following just breaks down under real conditions. I already talked about that. You know, telling it as an example to use ChatGPT Atlas and computer use, it'll open it and then use Chrome on its own. And also the permission prompts are required for every new app, in every from what I've seen and in my own testing, every single new run.
Jordan Wilson [00:18:31]:
Right? So even setting something up on a schedule, you might think, oh my gosh. Computer use on a schedule. I can run this every single night and, you know, do these little tasks. Well, yeah. But at least for me, at the beginning of each run, I have to authorize it each time. So, yeah, you can do that via the mobile app on dispatch, but it kinda takes away this concept of being able to, proactively automate and schedule computer use. Right? It anthropic in the app says you just have to do it once, and then it'll work. And at least for me, it doesn't.
Jordan Wilson [00:19:00]:
It's permission prompting each and every time. I can't wait for the equivalent of the dangerously skipped permissions. Right? If you use quad code, you probably know that. I can't wait for the equivalent of this, for computer control or computer use. They didn't even name this thing. Right? Because, technically, computer use, no one talked about this either. It's an another limitation, I guess. If you're going to look up computer use, you're like, oh, let me find some information on this.
Jordan Wilson [00:19:25]:
Good luck. It's not on their website. You gotta go find their tweet. And, actually, they came out with a, quote, unquote, computer use prod, product, I believe, in October 2024. Right? So it was essentially a even slower version of this. You had to run it via Docker. Right, and it didn't work at all. So, yeah, they didn't even really give this thing an official name.
Jordan Wilson [00:19:49]:
Most people are kind of calling it computer use, but technically, Anthropic came out with a computer use product in October 2024, and they don't even have this on their website. So alright. What could go wrong? Let's go live. Alright. And this is, for our livestream audience. This might seem awkward. I'm gonna go ahead and cut off my camera. Alright? Because I'm going to use my phone because that's what I record on, and I'm going to, run some of these examples on dispatch.
Jordan Wilson [00:20:17]:
So I'm gonna pull up and share my iPhone screen, as we kick this off. So give me a second here. You probably gonna, go blank here for a brief second. All right. Hopefully, hopefully we do this right. Live stream audience. Let me know. You should see my computer screen.
Jordan Wilson [00:20:37]:
All right. And let's go ahead and. Well, let me first describe what we have going on. So I am mirroring, this is a live, although it's being a little, finicky here. There we go. You'll see this is the live view of my iPhone. I have the, the iPhone Claude app open, and I have the, Claude Mac app open and I'm sharing my entire screen. So you have my Mac app, here on the left, you have my iPhone app here on the right in my desktop running in the background.
Jordan Wilson [00:21:16]:
So, we'll see. Hopefully this works. I might have to drag some things into view, but, we'll see. So I just sent a prompt that I already had kind of saved in because you didn't want to. I already said I had fat fingers and I hate my phone. So, I'm not going to try to type this out. But I am doing something fairly simple here and hopefully, this works. Okay.
Jordan Wilson [00:21:39]:
So I said for this task, use the computer use function, open codex on my computer. Right? So I'm having Claude use codex. This should be fun. All right. And then I'm saying, go. All right. Here we go again. All right.
Jordan Wilson [00:21:55]:
It's asking me to, allow for this session, even though I've already allowed this multiple times, I'm going to go ahead and click to allow. Alright. There we go. Okay. So a lot of things just jumped on my screen. Don't worry. I'm gonna tell you what I said. So I said for this task, use the computer use function, Open codex on my computer.
Jordan Wilson [00:22:19]:
Go to the build Mac OS Pomodoro timer entry on the left hand side and click into it. At the bottom of this page, find the bash command to run the program in the terminal. Run the command to open the app, then click around the user interface for a few screens. Then send a message to codex with one specific piece of feedback or improvement based on your findings. And then I also said, I was trying to run this task a couple times, and then it kinda refused. It's like, I just did this. Right? And I'm like, I know I just did it. Please do it again.
Jordan Wilson [00:22:53]:
Alright. So that's all we have doing. So now I'm gonna try not to really move, my mouse, but, hopefully, we'll be able to see, what's going on. And I won't have to jump in because like I said, it does sometimes run on other screens, because I do have a, kind of two screen setup right here. Okay, cool. So it did this and it ran this app. So it did the first part. Well, so, side note secret, this little app that I've been working on a productivity app.
Jordan Wilson [00:23:27]:
I wish I had more time. Right? But, I tell you all a lot of times I always have, like, codex and cloud code open, and there's, like, three or four apps that I've been working on, apps that I've used myself. But I think ultimately, probably a lot of people would want to use. You know, maybe maybe I'll, share more on that in our community. So if you're in our inner circle community, maybe go bug me about that and be like, hey. What's up with that? So, anyways, Claude Code is doing what it's supposed to. It, launched this app. So it opened first codex.
Jordan Wilson [00:24:00]:
It found this command to open this app that's still in development. It's clicking around the actual screens. Right? Which is really cool. So this was not, an option to do right like a week ago before this came out and you will kind of see for our live stream audience. You can kind of see a faint, orange glowing highlight around it. So that tells the user that it is using computer use and, well, also, it sees everything that's on the computer. So if I had anything else on that screen, it is going to see that. So, again, when you talk about privacy and security implications, that's why.
Jordan Wilson [00:24:36]:
So if you, you you know, have your text messages up or, you know, your bank information in another window on the same screen and you're sharing an entire screen, it will see all of that. All right. So now, one thing I'm not the biggest fan of, I'm gonna actually make that a little bigger there. Okay. AI moves too fast to follow, but you're expected to keep up. Otherwise, your career or company might lag behind while AI native competitors leap ahead. But you don't have ten hours a day to understand it all. That's what I do for you.
Jordan Wilson [00:25:18]:
But after 700 plus episodes of Everyday AI, the most common questions I get is, where do I start? That's why we created the Start Here series, an ongoing podcast series of more than a dozen episodes you can listen to in order. It covers the AI basics for beginners and sharpens the skills of AI champions pushing their companies forward. In the ongoing series, we explain complex trends in simple language that you can turn into action. There's three ways to jump in. Number one, go scroll back to the first one in episode six ninety one. Number two, tap the link in your show notes at any time for the start here series, or you can just go to starthereseries.com, which also gives you free access to our inner circle community where you can connect with other business leaders doing the same. The Start Here series will slow down the pace of AI so you can get ahead. Cool.
Jordan Wilson [00:26:16]:
So it did it correctly. Claude Code went in, opened codex, launched this app, clicked around, find some things that maybe didn't work correctly, and now it's typing a message to codex. Unfortunately, I think I kind of okay. There we go. I was gonna say I had to move the mouse to try to resize, the Claude code window so we could see what's actually going on. But it did it. So it said, let me kind of click over here. So it gave a message to codex, and it said in it said UI feedback.
Jordan Wilson [00:26:52]:
In the next tab, the accents, red for overdue, orange for due today, neutral for upcoming, so users can instantly triage their query at a glance without reading each label. So there you see right there. Alright? This actually worked. Right? Like, this, I knew was gonna be consistent. It was at least gonna be a decent enough demo that you could see and understand, you know, some of the utility, from this new feature from Anthropic. There's a lot of other things, and I don't know why. Claude computer use is just really bad at browsing the Internet. So I I I try to in Claude code with the, or sorry.
Jordan Wilson [00:27:36]:
I I tried it in Google Chrome, with the Chrome extension. Not very good. Yes. It does just take screenshots of everything, but the browser use, you know, on all benchmarks, the anthropic models usually do fairly well. So for whatever reason, it really struggles. I don't know if this is again a bug. Again, keep in mind, this is a research preview. So if you wanted to do a bunch of agentic browsing, probably not the best use case right now.
Jordan Wilson [00:28:06]:
Right? If and this might sound crazy. Something like Chad GPT's agent mode, which hasn't been touched or updated, it seems like in a year and isn't that good. It's actually pretty good for doing things like that, probably better, or using, like, what I use. Codecs, has a great playwright and interactive playwright integration, which is essentially just a dedicated Chrome browser. Codecs is amazing in that. So for whatever reason, Cloud Code not great, at instruction following in the new computer use. It's not great at using the browser. But for other things, opening other apps, like, what I just did here was actually pretty good.
Jordan Wilson [00:28:44]:
Right? It worked. And it came back. It said you know, it reported back to me, in the kind of dispatch terminal here. It said, done. Here's what I did step by step. I opened codex. I found the the project you told me. I ran the batch command.
Jordan Wilson [00:29:01]:
You know, I clicked through the four different screens, and I sent this feedback to codex. So did a pretty good job, and you can kinda see in the background here, codex is is running and doing all the updates that Claude Code, talked about. Alright. So let's wrap this thing up and I'll answer the question. Is the new computer use that went mega viral, is it a fun party trick or is it a real agentic workforce multiplier right now? And well, the answer is today, it's a party trick, but I think the architecture points to something much, much bigger. Right? We talked about this, you know, earlier, this week, on the show. And I think I mentioned this briefly last week as well. I think this is the next layer of knowledge work.
Jordan Wilson [00:29:55]:
I think this actually will be pretty big. And I do see this rolling out pretty quickly to Microsoft, right, which might be shocking. But, little do most people know, it seems like Microsoft is now going to have a dedicated, senior, I think is a vice president working on an OpenClaw integration. Right? So, Microsoft has actually been shipping a lot of things very quickly. They shipped a version of Cowork. Right? In like, very similar to anthropics, co work. Microsoft has co work. Microsoft has tasks.
Jordan Wilson [00:30:32]:
So I wouldn't be surprised for something like this, a, fully, functioning, computer use product like this for Microsoft. And when that happens, I do assume, that OpenAI and Google wouldn't be too far behind. Right? Google already has this built into their, their browser with the Mariner project. You know, OpenAI obviously has this on a virtual browser, virtual sandbox via ChatGPT agent mode. So I do assume that when OpenAI goes to the super app, this will be a future functionality. I assume that Microsoft and Google will follow suit as well. So I do see this even though it's buggy. Yes.
Jordan Wilson [00:31:14]:
It's a research preview. Right now, it's a party trick, but I also firmly believe that this is the future of work. I don't think this is the end medium. Right? But I I I think I mentioned this before. I think there is a a a brief period of maybe a half year to a couple of years where this is going to be the next interface. Right? Until, more and more websites, more and more apps do a little bit better at supporting agentic protocols, you know, a to a, you know, MCP. There's all these, you know, ACPs, these, agentic checkout protocols, all these different things. Until, you know, the essentially, the entire Internet, all SaaS products support those, This is going to be the new human duct tape.
Jordan Wilson [00:31:56]:
It is going to be the AI duct tape, and that is in a GENTIQ computer using layer. So even though it's not great right now, I still think it's promising. And I do know that, entropic will continue to ship updates because they've been shipping like crazy. So I do think that this will improve. So I hate being that guy that always says this is the worst it'll ever be, but this is the worst it'll ever be. So you should probably go in now and start using it and get familiar. And a couple pieces of advice for me, I've been using this a lot since it first came out. Plan to run any task at least two to three times, and you have to very closely audit the results even for very simple workflows.
Jordan Wilson [00:32:35]:
Right? So Anthropic, you know, recommended to start with, you know, simple tasks, and I agree. But you really do have to run them multiple times, look at the results, and even ask Claude like, Hey, this didn't go right. How should we reframe this? So it actually works consistently. And I think this is going to be. The first, one of the first major pieces for an actual super agent that works everywhere. Because like I said, there's obviously great AI agents out there that can automate all of this without you having to look at anything. But I think the majority of enterprises right now, well, they're using different software or for whatever reason, you you know, having things locally and not in the cloud where, you know, they're not able to integrate things directly. So I do think a lot of what us humans are doing.
Jordan Wilson [00:33:25]:
Right? The copying and pasting between different AI systems, you know, exporting data, uploading it into a large language model, for evaluation. I think this is the next layer. It is the agentic computer using error, not error, although there's actually been a lot of errors, but layer. And I do think it's important for you as you put AI to work this Wednesday to get familiar with it. Even if anthropic Claude isn't your go to, might be worth signing up for a month to give it a go just to see how it works. Alright. That's it. I hope this was helpful.
Jordan Wilson [00:34:01]:
The review of anthropic Claude's viral new computer use feature. Hey. If this was helpful, let me know. Alright. If you wanna see more shows like this, we do these more hands on demos on Mondays on, or sorry, on Wednesdays. On Mondays, we bring you the AI news that matters. That's a recap of the biggest and most impactful, trends and stories and happenings of the week. On Fridays, we do feature Fridays going over some of the, sometimes looked over or glossed over, features and updates from large language models.
Jordan Wilson [00:34:32]:
And then on Tuesdays and Thursdays, we rotate shows. Right? So if this is helpful, please consider take thirty seconds, subscribe to the show, leave us a rating. I would really appreciate that. If you're listening here live on the livestream, please repost this, tag one of your friends on LinkedIn that needs to hear this. And then when you're done, go to youreverydayai.com. Sign up for the free daily newsletter. Thanks for tuning in. We hope to see you back tomorrow and every day for more everyday AI.
Jordan Wilson [00:34:58]:
Thanks, y'all.