앤트로픽의 AI 에이전트: 가상 지원으로 작업 자동화하기

### AI의 발전: 로우레벨 AI 에이전트를 통한 작업 자동화
* 🤖 이번 주에 Anthropic은 우리를 대신하여 작업을 수행할 수 있는 **낮은 수준의 AI 에이전트**를 도입하여 **AI 기능의 비약적인 발전**을 보여주었습니다.
* 💻 AI는 명령만 내리면 수동 입력 없이도 **YouTube 채널에서 인기 동영상을 찾고 데이터를 스프레드시트로 컴파일**하는 등 웹을 자율적으로 탐색할 수 있습니다.
* 🔧 이 기능을 설정하려면 **Docker를 설치**하고 **API 키를 구성**해야 하는데, 지나치게 복잡하지는 않지만 시작하려면 몇 가지 단계를 거쳐야 합니다.
* 🌐 설치 후 사용자는 가상 데스크톱 환경을 통해 AI와 상호 작용하여 **날씨 데이터 수집, 스프레드시트 채우기 등의 명령**을 실행할 수 있습니다.
* ⚠ 그러나 AI는 **요청에 대한 속도 제한**과 같은 한계에 직면하여 다단계 프로세스를 원활하게 완료하는 데 방해가 될 수 있습니다.
* 💳 유료 티어로 업그레이드하면 **요청 한도가 증가**하여 중단 없이 더 원활하게 작업하고 더 광범위한 작업을 수행할 수 있습니다.
* 🖌 그리기 작업의 어려움과 같은 일부 버그에도 불구하고 **다양한 디지털 작업을 자동화하는 AI의 능력**는 향후 AI 기술의 발전 가능성을 보여줍니다.
* 💡 AI 도구에 대해 더 자세히 알아보고 싶은 분들을 위해 **미래 도구와 같은 리소스에서 AI 분야의 최신 혁신과 튜토리얼**을 확인할 수 있습니다.

앤트로픽의 새로운 AI 에이전트는 명령을 사용하여 자율적으로 작업을 수행할 수 있어 인공지능 기능의 상당한 발전을 보여줍니다.

Anthropic launched something amazing this week, and we officially have low-level AI agents that can go and do work on our behalf for us now. So, check this out. I can give it a command like find the five most popular videos on Matt Wolf's YouTube channel and add the dates and title of each of them to a spreadsheet. Press enter and watch as the tool opens Firefox for me, navigates to YouTube for me, searches out Matt Wolf, clicks into my channel for me, clicks on the video tab, sorts it by most popular, opens up a spreadsheet for me, and then watch as one by one it adds each of the titles of my most popular videos and pulls in the time since it was released. It just scraped it straight from my YouTube channel and is creating the spreadsheet hands-free. It just went through that whole process and just did each of the steps until I have the spreadsheet that you see on the screen right now. I'm talking as it's doing it. I'm not doing anything. This is crazy that we're at this point with AI where

It can go and use tools and do work for us on our behalf by giving it prompts. You want to know how I did this? Let's break it down. All right, so let's go ahead and set up this new computer use feature. Now, it's not the most intuitive process to set it up. Like, you can't just use it inside of CLA right now. You have to install some stuff on your computer and set up your API key, and it's really not that complicated, but it's not as simple as just going to a simple website and being able to use the tools. So, the first thing you need to do in order to set this up is we're going to install Docker on our computer. So, if we head on over to docker.io, after about a minute or so, we have Docker installed. So, I'll go ahead and close out of this. Now that Docker's installed, let's just go ahead and open it up in the setup process. We're just going to go ahead and use the recommended settings, click finish, and

Now I've got Docker set up on my computer. So now that we've got Docker installed and running, we're going to click on this little terminal button here. This says a terminal directly in Docker desktop. Click that. Go ahead and enable your terminal here. What we need now is a cloud API key. To get that, we're going to go to console.anthropic.com, and you're going to need to log in. I logged in with my Google account. You're going to see a menu that looks like this. Go ahead and click get API keys and create an API key for the name. I'm just going to put test key. You can name it whatever you want. It doesn't really matter. It's just so you know what you created that key for. I'm actually going to delete this key after the video, so not a big deal. And I'm going to save it to my default workspace and click add. Then we'll get this API key here. I'll go ahead and copy that. I have this little snippet of code here. I'm going to put

this code in the description so you can easily copy and paste it on your computer. But we need to do something with this code first. So here you can see anthropic API key equals and then dollar sign anthropic API key. We're going to go ahead and select this whole section where it says dollar sign anthropic API key and we're going to paste in that API key that we just grabbed from the console here. Now we're going to copy this entire code that we just created again. It'll be in the description. And let's open up Docker again. I'm going to go ahead and click on this terminal button and then paste this code right into my terminal and hit enter. And you can see it's firing up this whole thing for us. It takes less than a minute. And when it's done, you'll see this little call to action that says open Local Host colon 8080 in your browser to begin. I just click on this. You'll see that it opens up Local Host 8080 and it also has

like a fake desktop over here on the right, and I can give it commands. So, you can see we've got like Excel, we've got like a terminal, Firefox, PDF viewer, calculator, like a paint app, and various things that it can use on this desktop. Now, I can't actually do anything on this desktop unless I click on toggle screen control. If I click on this, now I have access to actually use the desktop myself. But I don't want to use the desktop myself; I want the AI to use the desktop. Now, over on the left, we can give it commands, and it will do multi-step processes and use the various tools that are available. So, I can give it a command like find the weather for the next 3 days in San Diego and then create a spreadsheet that contains each date and the high and low temperatures. Let's see if it can actually run this process. You can see it says running agent, and it says, I'll help you get the weather information in San Diego and create a spreadsheet tool use computer. Now it's

opening up Firefox. I'm not doing any of this. It's opening up Firefox on its own. It took a screenshot to see that it was actually using Firefox. Then it typed in San Diego weather forecast for me, figured out the temperatures. Now it opened up an Excel spreadsheet. Again, I am not controlling the desktop at all. It's literally just taking screenshots every so often to figure out what it needs to do next. So now it's got a spreadsheet open, and it ran into its first error here. It says runtime error: number of request tokens has exceeded your per minute rate limit. So apparently, it has some limitations where if it needs to go through too many steps to get you to your end result, it falls apart. But let's go ahead and see what happens if I just type please continue the process. See, it says running agent again, and it seems like it's getting back on track. Now it just added the date column, the high temperature, and the low temperature column. I'm not doing any of this. Added the date, the high temperature.

now the low temperature for October 23rd. Now it's doing the same thing for October 24th. This is based on all of the information that it just found by searching the web. So it searched Firefox for weather in San Diego for the next 3 days, found that information, opened up a spreadsheet, created the spreadsheet for me, and then filled out all the details on the spreadsheet. Now it's actually saving the file as San Diego weather. If I scroll through, you can see each step: tool, use, computer input, action, type, text, date. You can see it added the date in this first column here. And then what it does is it takes the action, takes a screenshot to see the current state of the computer, takes the next action. See, you can see here it made the high temperature column and then took a screenshot to kind of confirm it did everything right. Then it added the third column for low temperature, took a screenshot to make sure everything was looking good, and it went through the process of doing the thing, taking a screenshot, doing the thing, taking a

screenshot to confirm that it's actually going through each of the steps properly, and our final result here is a spreadsheet that I can click around on, and it's got the date, the high temperature, and the low temperature. If I open up Firefox here, you can see this is where it grabbed the data from San Diego weather forecast, and it grabbed it for the next 3 days. Here, I'm going to go ahead and refresh and start over so my chat window on the left is clean. Let's tell it to find a photo of Matt Wolf, the YouTuber, download it, and then resize it so the height is 500 pixels. And let's go ahead and let it run that. Let's see if it will automatically do all these steps. My hands are off of it now, so it's looking at what's currently on the screen to figure out what it needs to do. It sees we got Firefox with the weather. It just did a search for Matt Wolf, YouTuber. Now it's clicking on the images tab to find some photos.

and once again, it looks like I ran into that same error code: rate limit error. The number of request tokens has exceeded your per-minute rate limit. I can also contact sales at Anthropics to discuss my options for a rate limit increase, but let's go ahead and do what I did last time and just tell it to please continue, and it picks up where it left off. You can see it's saving the image here. It took a screenshot to see the current state of things, and now it won't even continue even when I type please continue. So apparently, that's just too many steps for it to handle right now. But the big idea is we can see where this is all headed. I'm going to go ahead and refresh this again, get a whole new setup going. You can see if I go back to my Anthropics console here under my limits, you can see it's got five requests per minute on Cloud 3.5 Sonet, which is what it's actually using. So apparently, it's just sort of moving through those steps too fast to where I'm

hitting that limit. Okay, so what I just did was I entered some credit card details on the billing section so that I'm no longer on what they called the free tier. Now I'm on tier one, and it looks like it upgraded my requests per minute to 50 instead of five. So now I should be able to go back and use this agent feature without it sort of hitting those rate limits. So now let's try this again, hopefully with an increased rate limit. I'm going to use the prompt once again: find an image of Matt Wolf, the YouTuber, download it, and then resize it so that the height is 500 pixels. And let's let it do its thing, hands off. I'm not doing anything. It took a picture of the screen, says it's going to open up Firefox now. It's opened up Firefox, did a quick search for Matt Wolf, YouTuber, found a bunch of images, went and clicked to the images tab. Now this is a little bit slower than what you're seeing; I'm cutting out some of the gaps.

in the video to make it feel like it's moving along, but there is a little bit of a delay between each of these steps because what it's doing is it's looking at a screenshot to see the current state of the screen. And then what it is doing after it looks at the screenshot is it decides the next action. Like right now, it sees that it's on a file naming section, so it named the file it downloaded as Untitled JPEG. And now we can see it's using a command down here to try to resize the image that it just downloaded. So this is actually pretty insane. I want to test one last thing 'cause there's like a little paint program here. Let's tell it to open paint and draw a stick figure. Let's see if it can draw. I don't even know what's going to happen here. Okay, so it opened up a tool called color paint here. Now let's draw a stick figure. A head, a circle, vertical line for the body, arms, horizontal lines, legs, two diagonal lines.

so now it's looking at the screen and uh let's see does it actually draw the stuff or is it stuck in a loop seems to be stuck on the drawing step doesn't quite seem to be figuring out how to draw an actual picture right now there theoretically should be a circle already here but we're not seeing it I'm going to go ahead and stop this one it doesn't seem to be uh getting the point doesn't seem to be able to actually draw what it's supposed to be drawing but we can get the concept we can see what anthropic is going for with this we give it a command of what we want it to do and it will actually go browse Firefox for us it will open up a spreadsheet and enter data for us it will do research for us it will download files for us it will save files for us it will run commands in the terminal for us now this is running inside of a virtual computer this isn't actually running on my computer this is all virtual here so

anything that actually happens on this fake computer here isn't really affecting my real computer, but we can see all of the tools that they're demonstrating it can use. Now, pretty wild stuff. We're getting this much closer to actual AI agents. Now, this is technically an AI agent, but we've seen in this demo that it still leaves a little bit to be desired. It still runs into rate limit issues even when I upgrade it and have a limit of 50 requests per minute. It actually still hit my rate limit when I was trying some of my things. Also, it was trying to draw like a stick figure by going left click drag on these coordinates, but for whatever reason, nothing was actually happening inside of our paint program. But it did manage to complete my command where I asked it to look up the weather for me and then open up a spreadsheet and then enter all the details of the weather in the spreadsheet. It did manage to do a search for my face and downloaded an image and then resize that image when I asked it.

to draw a stick figure, it did go and actually install a new paint program to use, but then couldn't quite figure out how to draw the stick figure. It's pretty dang impressive where this is all headed. Again, still a little bit to be desired. There's still some bugs. This is still a really early demo, but nonetheless, really fun to play with, really cool to put it through its motions and see what it's capable of. I'm excited to dig in even more and play with it some more and figure out what sort of agentic tasks I can get it to do. But I wanted to make a super quick video for you, show off what it's capable of, what it's not quite good at yet from my own testing, and ideally make it super easy for you to go follow these exact instructions, install this agent on your computer, and put it through its motions, see what it can do. It's really fun. It's really cool to see a tool actually take actions for us now and use these various tools. Really excited to see this improve over time.

Hopefully, you enjoyed this video. Hopefully, you learned something at the very least. Hopefully, you could go and play around and run with your own AI agent yourself. Uh, if you're curious about the cost, I've been playing with it for a couple of minutes here. I did add $25 to it just to sort of increase my rate limit there. All of my messing around that I did just for this video cost me about a buck 50. So, it's still fairly inexpensive. If you were to build this into a software app and it was using it a lot, it would add up quickly. But just playing around with it, you're going to spend, you know, a couple bucks maybe. If you want to find out about a whole bunch of other cool AI tools, check out Future Tools, where I curate all of the cool tools I come across. Also, check out the AI news page. Any new big breaking AI news that comes out gets listed on this page pretty quickly after it happens. And there's a free newsletter where I'll send the coolest tools and

the most important news directly to your inbox. I'll also hook you up with the AI income database, which is a sort of list of cool ways to make money using various AI tools. Just kind of a fun thing that I put together, but that's free for you if you join the newsletter. Again, all free over at Future tools. And uh, that's what I got for you today. Hopefully, you enjoyed this. If you like videos like this, you want to stay in the loop with the AI news, you want the latest AI tutorials, you want to know how to use all these cool AI softwares that are coming out, make sure you like this video and you subscribe to this Channel, and I'll make sure more videos like this show up in your YouTube feed. Thank you once again for nerding out with me today. Really, really appreciate you. See you in the next one. Bye-bye.

How To Install and Use Claude's New AI Agent

AI, No-Code, Tech, Futurism - I'm a tech nerd and talk about tech nerd stuff

Matt Wolfe

로우레벨 AI 에이전트: 게임 체인저

Anthropic은 웹 사이트 탐색 및 스프레드시트 작성과 같은 작업을 대신 수행할 수 있는 저수준 AI 에이전트를 출시했습니다.

설정 과정에는 Docker를 설치하고 Anthropic에서 클라우드 API 키를 받는 과정이 포함됩니다.

설정이 완료되면 AI 에이전트는 핸즈프리로 명령을 실행할 수 있어 업무 자동화에 있어 AI의 강력한 기능을 보여줍니다.

이 과정은 제공된 코드에 Anthropic API 키를 붙여넣고 Docker 터미널에서 실행하는 것으로 시작됩니다.

그러면 다양한 도구가 포함된 가상 데스크톱이 실행되어 AI가 자율적으로 명령을 실행할 수 있습니다.

AI는 샌디에이고의 날씨를 찾고 스프레드시트를 만들려고 시도하지만 속도 제한 오류가 발생하지만 프로세스를 계속 진행하여 필요한 열을 스프레드시트에 성공적으로 추가함으로써 이 문제를 해결합니다.

컴퓨터 프로그램으로 작업 자동화하기

이 프로그램은 향후 3일간 샌디에이고의 날씨 데이터를 성공적으로 검색하여 스프레드시트를 만들고 세부 정보를 입력했습니다.

이 프로그램은 스크린샷을 사용하여 프로세스의 각 단계를 확인했습니다.

그런 다음 이 프로그램은 유튜버 Matt Wolf의 사진을 찾아서 다운로드하려고 시도했지만 속도 제한 오류가 발생하여 작업이 일시적으로 중단되었습니다.

요청 증가를 위해 티어 1로 업그레이드하기

이 사용자는 신용카드 정보를 입력하여 티어 1로 업그레이드하여 분당 요청 수를 5개에서 50개로 늘렸습니다.

이 사용자는 유튜버인 Matt Wolf의 이미지를 찾아서 다운로드하고 크기를 조정하라는 메시지를 표시하여 에이전트 기능을 테스트했습니다.

에이전트는 Firefox를 성공적으로 열고 이미지를 검색하여 다운로드한 후 크기 조정을 시도하여 가상 환경에서 작업을 수행할 수 있는 능력을 보여주었습니다.

AI 에이전트 데모는 인상적인 도구 사용법을 보여주지만 속도 제한과 특정 작업 실행에 어려움을 겪습니다.

날씨 조회, 스프레드시트 열기, 이미지 다운로드와 같은 명령은 성공적으로 완료했지만 막대 그림 그리기에는 어려움을 겪었습니다.

이 데모는 에이전트의 잠재력을 강조하고 개선의 여지를 남겨두어 실험하기에 흥미로운 도구입니다.

화자는 AI 에이전트와의 경험을 공유하며 몇 분 동안 사용하는 데 1.

50달러밖에 들지 않았다고 말합니다.

다른 AI 도구를 검색할 수 있도록 엄선된 리소스인 미래 도구와 함께 AI 뉴스 페이지 및 무료 뉴스레터를 제공합니다.

이 스피커는 시청자가 동영상을 좋아하고 채널을 구독하여 더 많은 AI 관련 콘텐츠를 보도록 권장합니다.