How Microsoft's AI VP automates everything with Warp | Marco Casalaina

0:00Marco Casalaina:Warp is pretty magical, but you can add to the magic and make it work more smoothly. You're talking about setting up little micro agents that do little tasks for you, either one-off ones like we saw in warp or recurring and triggered ones. And then this is making your life just easier. As soon as I started using it for certain things like managing Azure, giving Azure subscriptions and stuff like that, then I was hooked. I was like, man alive, this is a really capable tool. Until you start working with these agents, you don't really discover all the things that you can do with command lines. But I think once you start to test those, then it kind of opens up your mind to what is really possible.

0:44Welcome back to How I AI. I'm Claire Vaux, product leader and AI obsessive here on a mission to help you build better with these new tools. Today, I have Marco Casalania, VP of Core AI Products and AI Futurist at Microsoft. Marco is going to speed run through five AI use cases where microagents can reduce the friction of getting little tasks done, whether they're technical or not so technical. Let's get to it.

1:10Marco Casalaina:Meet Rovo, your AI teammate, connecting knowledge, people, and workflows so teams can work smarter and move faster. It helps people find answers, make decisions, and automate work securely and with context through search, chat, agents, and studio. Rovo runs on the teamwork graph, Atlassian's intelligent layer that unifies data across your first and third-party apps so no knowledge gets left behind. And you always get personalized AI insights from day one. And the best news, it's already built into Jira, Confluence, and Jira Service Management paid subscriptions, so the power of Rovo is already at your fingertips.

1:55Know the feeling when AI turns from tool to teammate? If you Rovo, you know. Discover Rovo, AI that knows your business, powered by Atlassian. Get started at rovo.com. That's R-O-V as in victory, O.com. Marco, thanks for joining How I AI. I am excited because we're going to see a tool, Warp, that we haven't yet seen on the podcast. And we're going to see you use it for maybe not its primary pitched use case, which is kind of agentic coding, but for some sort of more ancillary support use cases that you found to be really useful. So before we get into them, why have you hooked so deeply into Warp in particular?

2:46Marco Casalaina:I started using Warp ironically because one of our own teams here at Microsoft tuned me into it. It was our PowerShell team. And they were like, you should try this Warp thing. It automates PowerShell really well. And so I tried it. And as soon as I started using it for certain things like managing Azure and, you know, giving Azure subscriptions and stuff like that, then I was hooked. I was like, man alive, this is a really capable tool. And if you're looking for the sexiest episode of How I AI, it is this, because we are going to show you how to manage Azure resources with AI, which actually I'm making a joke because I think it's so funny.

3:28but these are the kinds of things that if you are a software engineer or an engineering leader or just building something, you are spending so much time on DevOps, admin, configuration, IAM, all that kind of stuff takes all your time and you don't actually get to the fun part of coding with AI. So, you know, show us maybe that specific use case and why you think Warp was such a good fit for that and what the pain was before you had a tool like this.

3:57Marco Casalaina:Yeah, let's do this. So I was working with my colleague Govind the other day, and I needed to assign him access to a number of Azure resources. And, you know, you give them granular roles. So here I needed to give him Azure AI user and Azure AI project manager. And this was part of a big project that Govind and I are working on. Now, to do this, it's actually not that easy, to be honest, to do this in Azure. And especially if I do it with a web interface. There is a web portal where I can go in there, and for each individual role, I can go find the role and assign it to Govind, and then the next role and assign it to Govind.

4:35Marco Casalaina:It's not very efficient. For all the roles I needed to give Govind, I mean, this would have taken me an hour. So instead, I do stuff like this. This is my prompt. I say, you know, I found Govan's email address in here to begin with. And then I'm like, okay, now give him Azure AI user and Azure AI project manager on this subscription that I'm looking at. And here it does it, right? So it will call AZ. AZ is this command line interface. And this is Warp's superpower. Aside from being a coding agent, which as I know, you know, a lot of people use it for. And I mostly don't actually. I actually use it more like this.

5:11Marco Casalaina:whenever there's a command line interface, a CLI that can do something, warp is freaking great at that. And so it will call AZ repeatedly until it runs to the ground. Now here, I think it made a mistake somehow, whatever it was doing at AZ roll, this one, it kind of made a mistake here. And then it got right back to it and it did it. And it's like, okay, I'm done. And then I say, okay, actually, I need to give him contributor roll on the whole subscription. and it does that too. No problem. And so I use this for all kinds of stuff here, but for Azure administration and close your ears, Microsoft people, I have also used this to administer GCP, worked just as well with G Cloud, the G Cloud CLI.

5:58Marco Casalaina:So it's great at this stuff. I was going to say, if you have been victimized by AWS, Azure, or GCP, admin interfaces for assigning roles. This is exactly the kind of workflow you want to see. And a meta thing I want to call out for people, because I've worked in DevTools for quite some time. And one of the challenges as a product person and an engineer working on DevTools is exposing a GUI on these very complex, very interactive sets of permissions, capabilities, configurations, it's actually a really hard design problem. It's like a very hard, you know, front end design problem. And what I love about AI having access to CLI tools, APIs, MCPs, all these ways to more programmatically access these capabilities, is you can actually abstract away all of that front end stuff and just let a user kind of query the system and get what needs to get done.

7:04And so if you're on the other side of this, you're not the user, but you're the builder of something like Azure. This makes it so much simpler to expose a quote unquote, like user interface to someone like you who needs to get get a job done. And then I have to call out a second thing, which is you're also doing what what I would I used to sorry RIP stack overflow a little bit but you know I used to like google how do I kill all processes for Adobe and then like find the command line you know uh the command and then paste it in and then you know you get the error and you paste it back into search and you try to find it and what I love about these more agentic processes that have access to the terminal and the command line is you can just do that all in one interface.

7:54Marco Casalaina:Totally. Yeah, exactly. Now, I will tell you, though, that there is a trick to making this stuff work. I mean, warp is pretty magical, to be honest, but you can add to the magic and make it work more smoothly. And there's a couple of ways you could do that. I mean, if you think about what I did with AZ, really, if you look at my MCP servers, well, this one's off right now, but I do connect this to the Microsoft Docs MCP server when I'm doing like Azure administration, because sometimes, you know, in this case, I knew exactly what roles I wanted to give Govan. But there are times when I have no idea what role somebody needs to do something like I'll be like, give this person whatever role they need to use Azure Document Intelligence.

8:36Marco Casalaina:And like you figure it out. Right. But rather than leaving it to its own devices, I can do as I'm doing here, I can connect it to the Microsoft documentation MCP server, which is a pretty good MCP server. And then it'll go look it up. And that makes it work much better. Another piece of this, and we'll see it again in a moment, is the rules. So now, originally, like out of the box, when I tell it, so I give it these rules. and so like if I'm giving rules on a resource group, roles on a resource group, I should say, I do need to activate my owner access first. So this is one of the common problems that I have is that I have not activated my owner access, which is like a hurdle I have to go through.

9:19Marco Casalaina:And so I make Warp remind me and Warp will be like, so hey, did you activate your owner access before I start doing this? Because otherwise it's going to fail. So you can give it these rules and MCP servers that kind of help it along and help it use this stuff. Of course, that's useful for coding as well. But, you know, I find it super useful for these kinds of things that I use it for. Well, and what I will call out, and this is no shade, but this is not the most sophisticated prompting I've ever seen in my life. It's just like, hey, when you're trying to do this, remind me to do that. If you're waiting on me, like pop open a browser, like wait for me to do the thing and then always use the CLI tool.

10:03And so I actually love looking at your prompting here because it's very conversational. And I think people get like wrapped around the axle on like my prompts have to be in a specific format and have to be these like very gracefully designed things. And honestly, for most stuff, just taking the step of writing like two or three steps that you need a system to follow are what makes things more effective. And then speaking of kind of like step by step processes, one of the other things I love about what you're doing with warp is you're just taking, again, I think these things that you could do in a UI and they would be annoying to do and not fun.

10:46and just having a smart technical agent do them on your behalf. So you want to walk us through how you did that with your daughter's homework?

10:56Marco Casalaina:Yeah, let's talk about that. My daughter is studying for a math test right now, and it's tomorrow. And her teacher gave her a practice test, and I decided to scan that in because sometimes what I'll do is she'll take the test, but I'll take the practice test, I'll scan it in, and then I feed it to ChatTPT, and I'm like, make me variance of these problems. And it will do it like inequalities and things like that. It'll make other inequalities that are similar, but different so that she gets a little bit more practice on these things. So I needed to scan this in. Now, my scanner has a feeder and so it sucks in the pages, but this was a two sided practice test.

11:36Marco Casalaina:And so I needed to scan the odd pages. Then I need to scan the even pages and I needed to put it together. So what do I do? I go to warp here and I say it exactly. Scan the documents from the feeder and save it to this directory as this file name. And it does. It totally does do that. And so it figures that out and done. It's done. Wait, can I pause you really quickly? Did this activate the scanner? Absolutely it does. What? Okay. Yeah. If I were home right now, which I'm in the office right now, but if I were home right now, I would do this again and you would see my scanner would wake right up and start scanning.

12:10So you didn't even have to press the little button and pop open the thing and you just, okay.

12:14Marco Casalaina:All I had to do was put the sheets in the feeder. And as soon as I typed this thing in, boom, there it goes. So for, for the youths listening and watching this show, as a, as a parent, you spend a lot of time with a scanner and a printer. I don't even know if you all know what that is, but this is, this is a peak, peak efficiency. If efficiency for me is being able to just remotely start the scanner from my AI coding tool. And I'm not going to lie. I scan her birthday cards and the Valentine's Day cards and things like that. I am deathly afraid that one day there will be a fire and all of these historical documents that she's made will be destroyed.

12:54Marco Casalaina:And so I scan those in too because it's easy, right? It only takes me a second. And I type this in here and now there it is. Okay. So you scanned one side. I scanned one side. Now, warp is kind of bimodal, so it works both as an agent and as a not agent. So then I just pressed the up key. So you see the second time I did this, it had generated this command line here. When I pressed up, it just would go straight to this command line. And so I'm like, well, I'm basically doing the same thing again, but now I'm doing the even pages. So I just changed odd to even over here. And it just did that. So it didn't even need to call the large language model to do this.

13:30Marco Casalaina:But then I wanted to put it together. So I said, now put together the odd pages and the even pages and just make the math practice test out of it. And so now it wrote some Python to do this, actually. I guess it installed PyPDF2 and it wrote a little Python file and ran it and then it removed it. So, I mean, in a sense, I am using here warp as a coding agent, ironically, sideways. But, you know, it did it. I mean, in the end, absolutely, it most certainly did create a unified thing. Now, I have to give these AI coding tools and IDEs, et cetera, a little product feedback, which is I want like a time to task, little timer here.

14:19And I want it to say, hey, you did this in 112 seconds and give me like confetti. because again this is one of those things I just think about how you would have done it before which is you would have walked over to your scanner which I have one right here I would pull up the terrible native scanner software you know if you know it if you know you know and I would have like scanned the thing and downloaded the pdf and then flipped it and scanned the thing and download the pdf and then like found some pdf opener and then like dragged and dropped the pages in the right the right order and then saved that file and it would have been so, so, so annoying.

14:56And instead you get to sit where we all want to live right now, which is in the terminal, in the terminal in dark mode, and just ask it to do this thing. And I think something that, you know, maybe less technical people don't appreciate is if you look at this and you look at these commands, Warp or Generating, a lot of stuff on your computer you're able to do programmatically. Right. And until you start working with these agents, you don't really discover all the things that you can do with command lines. But I think once you start to test those, then it kind of opens up your mind to what is really possible.

15:34Marco Casalaina:Well, what's interesting is it caused me to look for ways that I could do things with a command line. So this scanner thing, this wasn't magical. There is, as far as I know, not a way in Windows to CLI invoke a scanner unless you install NAPS 2. So if you look at my rule for this, and what I said is, when you're in Windows and I tell you to scan something, use NAPS 2 to do it. And I gave it the location of where I installed this thing called NAPS 2, which is this scanner CLI. It's an open source scanner CLI for Windows. So I installed this. I made a rule for it, knowing that I scan things so frequently that this would save me a vast amount of time.

16:15Marco Casalaina:So, you know, again, like it's magic, but it's not entirely magic. You do have to do a little preparation for this trick to work. This episode is brought to you by Lovable. If you've ever had an idea for an app but didn't know where to start, Lovable is for you. Lovable lets you build working apps and websites by simply chatting with AI. Then you can customize it, add automations, and deploy it to a live domain. It's perfect for marketers spinning up tools, product managers prototyping new ideas, or founders launching their next business. Unlike no-code tools, Lovable isn't about static pages.

16:51It builds full apps with real functionality. And it's fast. What used to take weeks, months, or even years, you can now do over the weekend. So if you've been sitting on an idea, now's the time to bring it to life. Get started for free at lovable.dev. That's lovable.dev. I am just laughing because, again, I think my pitch at the beginning of this podcast, which is this is the most glamorous episode of How I AI, which is we're going to do Azure role, role assignment. We're going to do drivers for your scanner that can be run via the CLI. And then you have one more use case, which is as somebody who does a podcast and works with a lot of videos is really useful that I thought maybe you could walk us through.

17:40Marco Casalaina:This one is for you. So I record a lot of videos myself, actually, and I have my little YouTube channel, although I can't say I have a podcast. It's not as regular as yours. But yesterday I used the game bar thing, the Xbox game bar thing to record a video off my screen. I said, maybe I'll try this, see how it goes. Well, for 10 minutes of video, this thing recorded a 1.7 gigabyte file. I don't know what it was doing, but I mean, it was insane. And I was recording this new tool that we're working on called Opal. Anyway, so I was like, what is up with this to warp? As you can see in my prompt here, I say, why is this file so big?

18:20Marco Casalaina:Use FFmpeg to reencode it, still keeping it at 1080p because I didn't want it to like nestify the resolution and make it more normal size. So FFmpeg is a CLI that you can use to edit videos. And I use this all the time. I use it to strip audio off of videos. One day, my coworker sent me a video where like from 7 seconds to 17 seconds, it suddenly went really quiet and then it went back to normal. So I said to warp, I'm like, use FFmpeg to like make the sound 300 % from 7 seconds to 17 seconds. And it totally does. But here, it looks at the file and it's like, okay, the video is 1.7 gigs because it has a huge bit rate and it's at a huge resolution for some reason.

19:03Marco Casalaina:And then it followed my instructions. It ran FFmpeg with whatever the switches were to re-encode this thing. And it did re-encode this down to 13 megabytes, which is what you would expect for like a 10 minute video of a screen share. Thank you. Yeah. And so again, like, I think this is one of those things that in an alternative world, somebody would have like gone to search and say like, video compression software, try to open something and like export and compress and figure this out. And instead, in just a couple seconds, you can use this more technical tool and get a lot of stuff done and also sort of understand the root cause.

19:44You know, another thing that I think people don't really appreciate about AI enough, and we had an episode with a producer from Ken Burns' documentary production agency is files are very rich with information and giving an agent access to a file you can tell a lot about that file and then if you layer on an llm you can tell a whole a whole lot about that file and so i do think file manipulation is a real underappreciated use like we do so much file generation but i actually think file manipulation is a really underappreciated use case of AI. Right.

20:24Marco Casalaina:Now, if you think about what I'm really doing with Warp, the way that I'm using Warp, I characterize it in a certain way. I call this an ad hoc agent because effectively each one of these things that I'm doing, you know, when I'm assigning the Azure roles or when I'm scanning the stuff or when I'm doing stuff with the videos, I'm kind of creating a little mini agent, an unnamed agent on the fly to do something for me. And that's becoming a trend. Like this is a thing that's starting to happen, not just in work, but in lots of different types of general purpose agents. Yeah. And what I would say is also what I love about AI and what I would recommend to people with AI is like, just get used to ephemeral stuff.

21:06Like just toss it. Like if you ever need to compress a video again, don't save this script. Don't like just just come back and do it again, probably with a better model at some point. And it's going to be just as cheap and just as easy. And so I think a lot of people get stuck in their head about like, oh, how do I make this a product or how do I get this production? It's like, don't get it to production. Just next time, do it over and do it over again. Maybe save a rule so you're not rediscovering the steps. Right. But like you don't need to build a whole a whole thing here.

21:36Marco Casalaina:And that's that's exactly the right idea. Right. So, you know, for example, like it happened once that it tried to call NAPS2 the scanner thingy and failed because it couldn't figure out what the path to NAPS2 was. And so that's why I made that rule. That's like, when I tell you to scan, here's where NAPS2 is. When I tell you to scan from the feeder, this is the switch to scan from the feeder instead of the flatbed scanner thing. And now that it has that rule, it has never done it wrong since then. It does it right every single time, even though I'm scanning to a different directory, a different file, maybe a different format.

22:08Marco Casalaina:It does it right every single time. You know, I'm not saying I love AI more than humans, but sometimes it would be really, really nice to be able to get that consistency out of the people around me. You know, perhaps my children who are not loaded with rules and context all the time and consistent output. No. Well, let's switch over to maybe some less technical use cases, but ones I think are really interesting. Again, thinking about ad hoc agents and workflows, how you're using sort of more administrative task-based workflow-based things to kind of be prepped for the work you need to do in a day.

22:51Marco Casalaina:Yeah, well, I mean, here I am in M365 Copilot. So this is Microsoft's general purpose agent for business. And a lot of people think of it like this. Like, I can ask it a question. When am I doing how? IAI. And it shows it here. Here you go. It knows my calendar. And that's cool. But what's happening now is that this and many general purpose agents like it are becoming agent builders. The line between consuming an agent and building an agent is blurring. So this is the new workflows agent. And this is an agent that builds an agent. So I'm going to kick this thing off. And what I said here is when I get an email from Clairvaux requesting a meeting at a certain time, check my calendar.

23:32Marco Casalaina:If that time is free, send her a 30-minute meeting invite for that time, and it will start to build this agent. Now, for the sake of time, I actually ran this in advance here, so I can show you what it will build in a second here. And what it has built is an agent. It's a triggered agent. It's an email-triggered agent. And so an email comes from you and it will extract the time from it. It knows enough to extract it in ISO 8601 format, which is the format that the API takes with the Outlook API. It will check my calendar and it'll create the meeting invite. And if I save this thing, this becomes a triggered agent that is now associated with my Outlook.

24:19Marco Casalaina:So if you send me an email and you're requesting a meeting, you're going to get an invite from me if I'm free. I'm not going to abuse it, but I do. I do love it. What I would say is really interesting here is the ability to set up synchronous response to asynchronous requests, meaning, you know, probably when I email you, you are busy. You are in a meeting. You do not have, I mean, I'm, I'm, I'm projecting now, but like, you don't have the time to look at your calendar and say, does this time work for me or not? But you know, when you have five minutes, you know, like, oh, I'm supposed to meet with Claire and she needs to be at the top of my queue because we want to get this podcast done.

25:01And so I'm going to set up the system. So as soon as she's ready, I'm ready. And I think that's a really nice flow. Again, I call this like burning down your anti-to-do list, which is if you can get yourself out of the critical path of doing a task and get AI into that path. Instead, you can be highly responsive and not drop stuff, which I think is really useful. And I will say we got this thing scheduled quite well. So if you ran this on me, it was really good.

25:33Marco Casalaina:Well, you know, you're a priority, Claire. He knows how to flatter the ghost. Oh my God. You know how to flatter the host. It's that Halloween podcast we did. That's right. That was such a fun podcast that you're still feeling the ghosts of it today. I am. Okay. And then, so this is sort of a more kind of reactive style agent that you built. What about a sort of like more Cron-based, like timeline-based one? Because you showed me how to do this in ChatGPT as well. Let's do this. Yeah. So like, what if you don't have M365 co-pilot? Now, this same kind of function is showing up in the consumer general purpose agents also.