GPT-6 Astra Computer Use: Fünf Praxisbeispiele im Test

One of the biggest upgrades with GPT-6's Astra is how well it can use your computer. I've been having it do everything from testing desktop applications that I've already built, to working through video edits, to even controlling applications through my phone from my computer. Mastering how to use this one feature will allow you to unlock all kinds of workflows, especially ones that don't have things like connectors or APIs. So in this video, I'm going to walk you through the five different ways that I've been using it, so by the end, you'll be able to adapt them to your day-to-day workflows. Let's get into it. Now, computer use on its own isn't a brand new feature. It's existed for the past year and I've covered it multiple times on this channel. But GPT-6, especially if you put it on something like high mode, takes it to the next level. Now, like I said, I'm going to walk through these five different use cases, and start off with a very familiar workflow. So when it comes to doing something like booking a flight, this is something that we've been able to do for more than a year using a basic version of computer use. But with the firepower of Astra, not only can we do it once, but we can build our own reusable workflow and create our own command line interface for a platform that doesn't have an API, or a programmatic way to look for things like flight deals. I personally have a keen interest in this, because back in high school, I had a small business where I would give access to a series of Python scripts where I would use them to find flash deals. Back in the day, this used to be rocket science. But now, you can do it with just a simple prompt. So not only are we going to execute a search and look for the best deal, but we're going to make it so that we can keep doing this whatever we want, and have programmatic access that you can make available to your agents. To run this, we're going to combine computer use with Codex's internal browser, and send over the following prompt: Build a reusable flight-search CLI for YUL to SYD, round trip December 20-25, 2026, two adults, in CAD. I want to be able to re-use it to setup a scheduled task to look for steep price decreases on business class flights to my favourite locations. First explore Google Flights in Codex's internal browser. Use a supported browser execution route and explain whether the CLI runs alone or requires Codex. Return price, airline, stops, duration, links, and retrieval time as JSON and CSV. Test a second date pair to prove reuse. Search only, do not book. To make sure it actually uses its internal browser, I like to explicitly call it out, because otherwise once in a while, it might use Codex and Chrome, and actually take advantage of your existing browsers, versus using its own. So if we send this over, you're going to see how quick it is, especially if you go on fast mode, where this is going to go almost instantly compared to where we were last year. It's going to find, ideally, the best deal, and start to reverse engineer how it got there, so we can do this more programmatically in the future. And after it executes the search and looks for the best deal, it starts looking at its own docs to see how it can build a vanilla command line interface that we can use from any agent with any model. So one of the many advantages of doing something like this, is that you can have Astra do the hard work the first time, and then you can execute this workflow with a lower-level model. Now after a few minutes, we get the first version of the CLI. But it seems to be over-indexed on this specific route. So then I ask it to do the following: Is it optimized to work with any combination and is it smart enough to test things like multi-city travel, etc.? Tap into all of your hidden knowledge on getting the best deal and execute any research if needed. Now if you don't know what that is, sometimes you can get a cheaper flight option on a specific route by going from one city one way to the next until you finally get to your destination. So not only do I want it to be more resourceful out of the box, but I also want it to tap into its hidden knowledge on training data, because I'm sure it's gone through tons of blogs and snapshots on how to use points, different types of points, and go to certain types of airports to get the best bang for your buck. Now it's testing different combinations of cities to see if there's any difference whatsoever in the price. After some more tweaks, we have a more robust version that's more flexible, can choose different depart and end dates, and structure a trip in different ways. But this is not actually helpful, unless this CLI is way faster than redoing this process every time using computer use. So the next step is making sure that we test a random flight request in a brand new set of sessions, one that's using the CLI and one that's just using the plain vanilla computer use. This ran for close to seven minutes and these were the results. For the first test, planning Toronto to Lisbon, we went from 77 seconds with a vanilla browser to 23 seconds with the CLI, which is almost a 70% reduction in time. With the second test, we went from Toronto to Seoul, from 50 seconds down to 25 seconds, so almost exactly 50% less time. With this example, I'm just scratching the surface. There are likely many softwares and tools that you use in your day-to-day that might not have a backend, an API, or an MCP, so doing something like this where you use computer use to create a command line interface, so that you can then use it with any agent you want, whether it's a local model or a lower-tier Codex model, this will create a brand new path of least resistance. Now the second use case is probably something that you haven't tried before. If you've ever created or worked on your own MCP, whether it's being used by only yourself, a small team, or an entire company, you know that the hardest thing to do is simulate testing. More specifically, what does it look like for someone to take your MCP server, hook it up to their Claude or Codex, spin it up, and use it for the first time? So what I'm about to show you is Codex setting up and installing an MCP server into Claude Code, and then opening Claude Code and testing it to see how smoothly it runs, how much time it takes, and basically maps out the entire user journey. As an example, this is one of the MCPs that I put together recently. It's basically a layer between my community and its posts and my comments, that someone can tap into and use in Claude and Codex seamlessly to ask questions about general knowledge related to community content, or things that might have popped up in prior conversations. So let's say you already had your own tool or MCP, and it was ready to be tested. You could send over a prompt like this where you can literally ask Codex to use computer use to open Claude Desktop because it already exists on your computer, so it has access to it. And then test the existing EA Community Brain MCP. Ask it to find a classroom lesson about building or testing MCPs, then give it's title, source link, and three useful points. Show the actual tool call and check one point against the source. Now I'm going to add one more line here. I'm going to say, have a few back-and-forths so we can make sure that the user journey in using this tool is as smooth as possible, and use Claude Code tab. And once we send it over, I added one detail, which is to use the Claude Code tab, and this should be enough for it to open the app, double check it's on the right tab, then check it has the MCP server, then initiate the conversation. So you can see right here, it's double-checking that it's not on the Claude Code tab. We'll just click on always allow here. Now it is, so it should be able to go and click and double-check that the connector is indeed connected, and send over this request. So the best part of this is not only is it monitoring the conversation, it's also clicking down on the drop downs to see exactly what's happening behind the scenes, so you can see all of the lessons that are being pulled, the exact sources, and everything in between. It then reads the response and sends a follow-up. And the coolest thing is, you can apply this to all kinds of scenarios, whether it's Claude Code, or if there's no specific plugin for Gemini for whatever reason, you can open up the Gemini Desktop app and have Codex use it and in a way, have a back-and-forth between them without ever using an API or any form of plugin. And 20 minutes later, it comes back after a thorough back-and-forth, to not only give me feedback about the MCP, but also if I kept scrolling, it has source code that it monitored that said, basically the way you structured this function and this function is not very efficient and causes delays and latency. And again, this process is not limited to MCPs. If you built a series of skills and you want to make sure that your skill that works in Codex works just as well in Claude, this would be another way you could do that. All right, so use case three is my favorite and it's likely to be the one that's the most applicable to your day-to-day workflows. Whether it's a desktop app or a mobile app that you're using something like TestFlight on your computer to go through and test the user journey, you can ask computer use to take over the app, battle test it, in the exact same way you saw me do it in Claude. So let's say you're testing a desktop app, or a mobile app on your computer. As long as you give it the guidance, not only can you test it, but once it finds bugs, assuming that it has access to that code, it can keep improving the app, testing it, until the goal is reached. So let's take this app that I put together, which is my recreation of GrokBot, where I've connected it to a series of agents and has access to things like my Codex subscription, open router, etc. And I've even added the ability to easily add multiple agents and have them work together and give them their own computer. But of every feature that I've built so far, the computer use is the weakest because it's the hardest to teach a completely vanilla agent. So what if we did computer useception, where we use Codex's computer use to audit this app's computer use and keep improving it until it's perfect. So let's say we send this prompt: Open Grokky, my GrokBot clone, and test its computer-use feature through the app. Watch Grokky browse, check one requirement against the page, and reopen the conversation to verify it persists. Send a request around Google flights like we did earlier, that will explicitly ask it to use its computer abilities. And then, just to avoid you and I having to wait hours, we just want it to send a message, monitor it, and come back. In actuality, I would want this to run for hours using maybe something like slash goal, and wake up to a series of bugs that were identified and fixed. So when we send the request, it spins up the app and right now it should enter the prompt, there it is, it says, use your computer-use abilities. It should automatically allow this and then it should monitor what's happening with the computer use. And it's smart enough to realize that there's a component in the app where you can spy on what's happening. And there you go, there is my vanilla agent trying to use Google Flights, and we are watching computer use, watching computer use, to see how well it does. Now because I have all of the actions and the tool calls very transparent, it can start to really understand what is taking too long, what am I over-engineering or under-engineering to make this as smooth as possible. So my app's computer use ended up failing for two reasons. One, I'm using an older model, which doesn't have as big of a context window, and I have it built a feature that helps compact the conversation. So Codex was able to test it out and see very quickly that it failed the test, and theoretically, if you told it that its goal, slash goal, was to make this computer use as amazing as possible, it would see that failure, adjust the code, then redeploy, then respin up the app, and go through that loop until it was better. Use case number four is automating your video editing workflow. And whether or not you're a content creator, you should be able to at least learn a few tricks with this workflow that you can apply to something else. Now in my case, after I finish recording these videos, I render them in 4K, and depending on the length and whatever else is running on my computer, that can take anywhere between 40 minutes to four hours. Now let's say I'm up late recording a video like I am this one, and I don't want to wait until 3:00, 4:00 a.m. for it to render, to then make some slight edits before I ship it off to my editing team. So with the help of computer use, I was able to build a workflow where all I have to do is click render and export the video, and then Codex can keep polling and double-checking every 30 minutes how far the video is and how close it is to finally rendering. Once it's done, it can open up Descript, then take the video, upload said video to Descript, wait until it's ready, then send a prompt into the actual chat where you can't actually do this programmatically, at least not that easily, then wait for the response, click through all the different frames to make sure there's no random scene where I'm coughing or just dying, which happens more than you think, and then it can export it, double-check the export, send it to a Google Drive folder and then take that link and share it with my team. Now these tools like Descript have a command line interface of their own, and up until now, it's been able to take me to 80%. But now with this computer use, you can take processes that have been 80% possible and bring it all the way to 100% with the few parts that need a human in the loop. Now naturally, I can't render a video while I record a video, so I'm going to send a simulated request to take an existing video and just show you what the process would look like for Descript. So I would say, use the render-to-descript-delivery skill in demo mode. Go to my videos folder and I have a video related to the current state of the AI market. I want you to take said video and execute this process as is. So when I send this over, this should open the Descript app, then upload said video, then once it's done, start going through the editing process. Once it's done uploading the video, it sends this mega-prompt on my behalf, waits for the response, sees all the tweaks that happened to the transcript on the left-hand side, then it physically clicks through the video frame by frame to see if I'm looking a little bit too shady like this and sees if there are areas for us to cut out based on let's say silences where there's a lack of audio in one section and I do this to make it as easy as possible for my editing team to take the raw video and add some polish. And once it audits the transcript and make sure that all of my infinite stutters have been removed, it clicks on export in 4K, waits for it to export, then once it physically sees it on screen completed, then it will go and create a Google Drive link using the Google Drive connector and take the process from there all the way to opening up my WhatsApp on my computer and sending a message with the link. The last case isn't something that you would typically use computer use for, but it's unbelievably effective. So I just finished setting up my brand new Instagram account, where I'll be making all kinds of new AI content. And by the way, if you want to follow me, follow the handle that you see on screen right now, I'll start being active very soon. But anyway, as an Instagram noob, there are a series of settings that I'm supposed to toggle on and toggle off to increase the likelihood that I do well on the platform. Instead of me having to go through it myself and spend time on my phone, I would rather Codex use computer use to do the same thing. So we can send a prompt like this that says, use the iPhone Mirroring app I opened to configure Instagram on my actual phone. And let's add some detail. I just set up a brand new account and I want to make sure that any time I upload a new video, that we have all the settings optimized for a new creator to make sure that my videos do as well as possible. I'm aware from some Instagram videos I've seen in the past that there are some hidden features or settings that I should toggle on or off. I want you to take control of my Instagram, not send any form of DMs, don't do anything destructive, but go and optimize all of the easy settings that you can. So if we send this over, this should be able to open the iPhone Mirroring app on my Mac, take control of it, take control of the phone which I found to be a huge surprise and click through that specific app. Now that I've sent this over, it should open the iPhone Mirroring app, go and search for the Instagram app, go through, double-check it's on the right account, then go through the settings. So it's connecting to my phone, it should go and click on the search bar right now and enter the Instagram app. There we go, it's clicking through, it spotted it without even searching for it. It's already in settings and activity, so it's double-checking that it's on the brand new right account, then it's going to go back, click through, and do what it needs to do to make sure that everything is optimized as possible. So it goes to the media quality tab where it toggled on the highest quality, and it's just going through what seems to be any core things in its training data that tells it, these are the core things to focus on. And as this continues to go through my settings, you can imagine that any one of your apps that doesn't have a desktop equivalent, you can go through, navigate, clean it up, check features, check user journeys, etc. So you could theoretically spin up a series of sub-agents to go do some deep research for how to optimize your Facebook profile or your X profile or whatever, and then have it go through and implement all the changes that it found in its research. And just like that, 5 minutes later, it goes through and gives me a full checklist on what it did, what it toggled on and off, and that we're ready to go. And that's pretty much it. So hopefully this gives you a full breakdown as to how powerful Astra's computer use can be and how you can apply it to 10X your existing workflows. If you want access to any of the prompts that I showed you, including the Google Flights CLI that you can use for finding flight deals or automating your next trip, I'll make all of them available completely for free down in the second link. And for those of you that want to go deeper and master things like Astra and get more exclusive content, I've already released a separate video for my community on exactly how I'm prompting and using Astra for a variety of use cases. If you want to master Claude, Codex, and everything in between and find a way to monetize your skills, check out the first link down below and maybe I'll see you inside. For the rest of you, if you found this video helpful and you found something that might help you today, I'd super appreciate a like on the video and a comment if you so choose. I'll see you in the next one.