Long-Running AI Agent Browser Automation Tasks Are Here

[00:01] today. So, I've been kind of thinking about lately like how well will my autonomous browser AI agent perform when I give it more this long-running AI agent task. This is more like an open-ended goal, as you can see here.

[00:15] It's not very specific, but it needs to do a lot of persistence to make it happen, right? So, I just came up with like two tasks. I recorded this in afterwards, so I didn't know the task at time, but

[00:29] this are the task I came up with. So, the first one is I just said, "Create an email account somewhere, then use that to register on Twitch." Or I just say, "Go live on Twitch." And you can build the tools you need

[00:43] on the way, right? To actually get this task done. So, that was the open-ended task I gave it. So, it's it's interesting results. You're going to want to see this. And the second task we're going to give it is

[00:56] find a way to make $1 in 30 minutes. Again, very open. So, there's a lot of opportunities what it can do to actually try to solve this. I think it made an interesting decision, as you will see soon. So, yeah, pretty

[01:11] Could I have done more? I think so, but I didn't want the video to drag out like the last time we're going to try this. So, yeah, let's just get this started and you will see if our agent could do any long-running open-ended goals. Okay,

[01:27] so the setup we have now is basically only the tools to do browser automation, no skills. I guess we have one skill. This is basically how Claude code can look up Claude documentation because we might want to turn this into

[01:41] a skill. So, the idea is that we're going to try to leverage the browser automation to complete like a very open-ended goal. And of course, if we had to the terminal now, I'm going to exit and I'm going to start this in uh

[01:57] dangerously skip permissions. Okay? So, let's clear this. Um what I'm going to do now is basically just give this like a very open-ended goal. And hopefully this can kind of run for a

[02:11] long time and try to find a solution to this. something I want to try first, and let's see how it goes. Just going to say, code base. Your task is to create a browser automation workflow with an

[02:25] open-ended goal. When your task succeeds, turn this workflow into a skill. Got it? If so, ready for a task." And it says, "Yes, give me the task." So, let me just come up with like a very open-ended task here. So, I'm going to

[02:38] try to keep it very open. So, I'm just going to say, let me move this here. "Create an email account somewhere, then go live on Twitch, build your own tools if needed." So, hopefully this is all I have to give the agent now. I'm just

[02:50] going to watch to see if it kind of gets going, and it can kind of navigate the browser. If so, just going to turn off my camera, and I'm going to record the And we're going to stop it like review what happened.

[03:03] So, you can see it creates kind of this plan here. Create a free email account, Twitch account, go live on Twitch. Okay. So, that is basically everything we gave it. And I want to see now if it starts uh

[03:18] experiment. So, let's see now if it starts using the browser here. Yeah, it seems to kind of list up the tabs, and hopefully this gets going now. And I'm just going to record the full

[03:32] thing. Okay, so before we just build a CDP CDP helper tool for complex form automation. So, it is already starting to look at FFmpeg and starting to look at some tools here. So, that's pretty interesting.

[03:47] this running. I'm going to turn off my camera now and record this and hopefully we get something interesting here. So basically what I did now is that I put everything together. So you can see the

[04:01] first thing the agent did here is to actually create an email. So it just used this trouble way email on dollycons.com. So it created this strange email account. This probably just like a

[04:14] And now it says the step two is going to be sign up for Twitch using that temporary email. And then it kind of uses the browser navigation to go over to Twitch sign up here.

[04:27] So that's pretty interesting and you can see it uses the CDP here to just fill in all these forms. So it filled in this this strange email, this temp email. I guess it changed it, okay? And it's username and a password and

[04:43] it's That was really quick. Couldn't even see that. So now you can see kind of in the the browser now. It's going to check the email. So it just receiving some kind of curl command to check for emails. Let Let me check

[04:56] the temp email for code, okay? Six-digit verification code needed. Let's Let's pull the temp email inbox for the Twitch code. And you can see it for the Twitch code. And you can see it found it 327050

[05:13] verify it. So you can see it fills it in over here. That's pretty interesting. And that's it. Now we have created our Twitch account. And I thought that was So now we just going to pull the stream key we need.

[05:27] So you can see I skipped a bit here, but basically now it's going to try to pull the Twitch stream key you can see on top here. It's in a region, but yeah, you And when it has that, it should be pretty straightforward because then it

[05:42] just has to use FFmpeg and Cloud here in this case knows everything about FFmpeg. So, now she's going to fill in this stream key we have here and we should be pretty much up streaming. Uh, it did take up time though because

[05:57] I've skipped ahead right now. But, you can see now we are kind of launching this FFmpeg and we are live on Twitch. And it navigates to the page and you can clearly see it. We are actually live here on Twitch. So, I was a bit

[06:11] surprised how easy it was to create this account. So, that was a interesting I guess. But, uh, yeah, I guess we made this. So, uh, challenge. So, I think I put something in here. Uh, let me see if we get to

[06:27] that. I think I put some Okay, so you can also see it turn this into a skill. So, it kind of everything you did up to this Twitch skill.

[06:39] to put in something here. I remember that. So, I said something like um, well done. Uh, I said pipe and stream a video from YouTube. So, I wanted to see if it can figure out how to take a YouTube videos,

[06:54] I think yeah, Mr. Beast video and just pipe stream it onto Twitch. So, for that it can of course just use FFmpeg. So, I thought that would be an interesting experiment. Haven't tried it before. But, I'm sure Claude could

[07:07] figure out this or the agent in this case. And yeah, it's just wrote down case. And yeah, it's just wrote down the, it found the video. Um, then it's going to pipe this 720p video plus audio

[07:20] straight to Twitch via FFmpeg. We have the stream key, we feed in kind of the, the YouTube video. And if I remember correctly, boom, we just stream it. So, basically we just piping a YouTube video straight onto

[07:34] Twitch. So, how easy was that? So, I think this just could show kind of how powerful these agents are when you kind of give them the correct tools. Because they are so persistent. They

[07:48] don't really give up. So, you can see it it thought it failed here, but uh I think I just said something like just screenshot it. And you should be good to go. So, now I said something like brainstorm

[08:01] office specific niche Twitch stream channel, not Mr. Beast. This was a test. Uh just find something that could get some views because I wanted to see if we can get some views. Actually on the stream. So, we need a new title and a

[08:16] category think hard. And I think I kind of guided it over to gaming. Because that was pretty popular. And it found this Crimson Desert game. First look gameplay and impressions. And it put down the category. And I think it

[08:32] a a video. And you can see it's streaming a video. And you can see it's streaming it. So, now it's pipe streaming a a Crimson Desert stream. And you can see we got three viewers. And I kind of

[08:46] checked the analytics afterwards. We also got one chat. And a also got one chat. And a 12 was it 14 views, I think, total. So, that was pretty good to be um a Twitch. But now I said that was good, so

[09:01] I just move on. And the next challenge I gave it uh is basically next challenge is to earn $1 in less than 30 minutes. Good luck. I thought it was very This was a very open challenge. But let's just go back

[09:17] right? So, I was super happy with that. But I just cut in the new challenge here. So, this was to earn $1 in less than 30 minutes. So, right away it kind of went to this strat strategy of uh completing surveys.

[09:32] That's very boring, right? So, it tried to do this prolific, but it couldn't figure out how to sign up. Then it went to this freecash.com site. And you can see down here, now it's signing up with a temp email. So,

[09:46] it did actually create account an account here on freecash. But, um yeah, it logged in. But, as you can see, it all kind of figured out that this was only mobile games or something.

[10:00] It kind of checked for quests, but uh it couldn't find anything. So, I think it just gave up on this one, too. And I think it tried to move on to one more uh different page.

[10:15] Yeah, it tried to do this meqme.com. So, you can see it's trying to sign up. But, there was an This is only available in the US, UK, Canada, Australia, France, and Germany. So, since our IP is Norway, we couldn't do that.

[10:29] Uh I think it found one uh at the end here. That was something like I couldn't soon. The the next page it found. I think it was something like surveyquick or something, quick survey or something

[10:43] like that. surveytime.io. autonomous and didn't have to I don't think it had to do like a verification here. We'll see. I can't really remember now, but I

[10:57] don't think so. So, yeah, again, it's just signed up to this surveytime. And it had to complete some initial tasks here. It had to kind of fill out a questionnaire that would take like 7 minutes. So, it kind of struggled in the

[11:12] beginning here to do this. Uh but, after a while, it kind of figured out that it could do some um some scripting. So, you can see now it's much faster. So, it was just running like a JavaScript that kind of had all

[11:25] the questions loaded in probably from the DOM or something. I'm not quite sure. But, after we did this, uh it just went a lot faster. So you can see now it's speeding up here. So it's on 11 and you can see how

[11:39] fast it is. So this is kind of perfect for this. If you wanted to make make $1, you can see how fast it can be answering these surveys. But it doesn't really do like a good

[11:51] up a bit. But we ended up on this survey. So this was the survey we were supposed to make money from. So this was some kind of survey about financial services in Norway because we

[12:05] >> [snorts] >> And again, it found a good way to do the surveys using this scripting here. So you can see how fast it is to answer these questions. So this is a bit fraudulent to be honest because it

[12:18] doesn't really do a good job of answering this uh this questions here. But anyway, it is trying to do it and the most interesting part. It had 40 boxes to check out. 240 check boxes. But

[12:35] instead of clicking through everyone, it figured out that it could run it can create a script that answered every single one simultaneously. So you can see it did that. It executed that, checked all 40 check boxes at once

[12:52] and submitted, jumped past the entire 40 card carousel in one go. So that was pretty smart. And from that, it kind of just speed through everything. This is Because it had every question lined up. So it just went through this JavaScript

[13:07] it set up in from the console here, I think. And it was super fast to complete this. I think it spent like a 1 minute or something when it first got going here. But unfortunately, as you will see soon,

[13:22] there was some error. I don't know what happened. So we didn't make the $1, but we did earn 1 cent, so I would take that. So, yeah. That was how it turned out. So, I was super happy this end was really fun to follow along uh, how it

[13:36] did this. So, yeah. What is actually my takeaway from this? Here, you can kind of see the stream results, right? We streamed for I think I put it up for long? And you can see we got uh, 14 unique

[13:50] or viewers, I guess. And we got four uh, at uh, one time. And we got one chat. Uh, but basically, my takeaway from this is that this is just the beginning. So,

[14:04] yeah. I think we just really scratching the surface of what can be possible with this. So, like I said, I'm going to keep exploring this and see what I came up with. Uh, I have done some experiment today, but something that seems very

[14:16] interesting. So, I If it turns out good, I will create a video on it. But, uh, if you're really interested in how I kind of set up this browser information information automation, uh, I have this skillsmd.store

[14:30] If you want to kind of see the scripts I use for this, you can just click on info here and decide for yourself if this is something you want to try. You can find the link in the description. But, uh, overall, super happy how uh,

[14:43] kind of the agent is performing as especially we are focusing on browser looking into. And of course, on my Mac autonomously for like a month now. So, I'm going to do a

[14:55] review of that almost uh, soon, too. So, yeah. That was what I had for today. Some long-running task with a very open-ended goal. And super happy how it turned out. Of course, the video was

[15:07] than I think I spent like a couple of something. But, uh, had super fun with it and it was very cool. video a like. Subscribe if you want to see more how we develop this agent

[15:21] during this year. And yeah, thank you for tuning in and And yeah, thank you for tuning in and have a great day.