196 | How to Safely Run Powerful AI Agents (Like Manus & GenSpark) with no RISK

0:00Hello, and welcome to another live episode of the Leveraging AI podcast, the podcast that shares practical ethical ways to leverage your AI to improve efficiency, grow your business, grow your business, and advance your career. This is Isar, Maytis, your host, and we have a fascinating session for you today. Why fascinating? Because we're going to dive into advanced AI agents and how to use them safely. And I'll explain, first of all, what the hell all of that means, and then we're going to dive into a presentation, then we're going to dive to the actual tools and the process itself. But on a high level, a new type of agents emerged a few months ago, Manus earlier this year, and then GenSpark shortly after.

0:33Manus is a Chinese company, GenSpark is an American company, I think, from California. And they do similar things. They're kind of like a generalized agents that can do more or less anything that your browser can do. So it gets access to everything you have access to in the browser, including tools, including browsing websites, including writing code, including creating web applications, and so on. And you can do a huge variety of things with it. And I was very curious on how to use it and how I can play with it and how far can I push the envelope. But I was really scared because I run everything in Chrome.

1:03Chrome is my entire universe. And if something gets corrupted in Chrome, I'm screwed. like literally everything that I'm running runs in Chrome. And in addition, Chrome saves all my passwords to everything on the planet. And so if it gets access to that, I don't know what it might get access to that I'm not interested in. Or if it changes any of the passwords for whatever reason, then I'm screwed, stuck outside of tools that I want to use and so on and so forth. So I had this issue with giving it access to my computer or giving it access to my passwords or everything in my Chrome or my Google account and all these things.

1:32And yet I really wanted to test these tools. And that led me through this quest that I'm going to share with you today in detail. Now, at the end of the episode, after I show you the process on how to put these tools in a box where they're confined, but you can still do everything you want, I will show you a few examples of what you can do with these incredible tools just to inspire you to try yourself. So let's get started. I will start by sharing my screen. Those of you who are joining us live, first of all, thank you so much for joining us live. Those of you who are not joining us live, you should join us live.

2:01We do this every Thursday at noon Eastern time. You can join us on LinkedIn, you can join us on Zoom, and then you can ask questions and be in the chat and so on. But those of you who are here in the chat, feel free to introduce yourselves to where you're from, why you're interested in agents, and what's your current experience so I know what we're dealing with. And that's it. And we're going to get started. I will say one more thing. This is going to be a more advanced session. We're going to talk about a lot of advanced stuff, but 100 % of the things I'm going to share with you today, I did for the first time ever when I did this, which was, I don't know, two or three months ago when I did this process.

2:31So everything that I'm sharing with you may sound sophisticated and advanced and complicated, but I knew zero of it when I got started. And within an hour later, I was able to run these agents without any external guidance. So hopefully with my guidance, you'll be able to do it in less than an hour and have access to this amazing new tool and capability. In addition, we teach basic courses. We teach courses that we've been teaching for the last over two years. So the first course was in April of 2023. We teach at least one course a month. Right now, I'm actually teaching two in parallel. Most of these courses are private.

3:03Organizations and group invite me to teach their people, their organizations, consortia, different membership organizations, and so on. So if you're in such organization, feel free to reach out to me on LinkedIn or somewhere, and I'll gladly share the details with you. But we do public courses that anybody can join once a quarter where we have the bandwidth to do that. As I mentioned, we're just finishing one course that started a few weeks ago, and then the next public course is going to start on August. So if you want to know how to join the August course, and you should if you haven't taken any formal AI education yet, it's literally going to change your life.

3:33It's going to change your career. And if you're in senior position, it could change your business. So go check out the link. We'll drop the link in the chat right now and we'll drop the link in the show notes if you're listening to this afterwards. But with that, let me share my screen and let's dive into the topic for which we are here today. In the next few years, AI technology will change our world dramatically. Whether you are a business executive trying to catapult your business forward, or just somebody who refuses to be left behind and want to advance your career, this is the show for you.

4:07I'm your host, Isar Maitis, a serial entrepreneur and an AI enthusiast. You'll hear invaluable practical tips from innovative business leaders, AI practitioners, and some of the brightest AI minds in our world today on how you can leverage AI in ethical ways to advance your career and grow your business.

4:30So we'll start with a general concept of what the hell are agents and how they're different than a chat. Because we, people who are listening to this podcast, I assume most of you are already using multiple chat tools, whether it's Gemini, Claw, ChatGPT, DeepSeek, et cetera, your Choose Your poison or maybe you're more advanced and using all of them or some of them depending on the specific use cases. And you're asking yourself, what the hell are agents and how they're different? And there's several different ways in which agents are different than chats. The main thing is autonomy. Agents have levels of autonomy.

5:01Some of them are fully autonomous. Some of them have levels of autonomy. But if you think about it, a chat only does what you tell it to do. You have to explain to it. Go and check the information on this website. Summarize the information, put it in the table in this and that format. Create a code to do this or that. You tell it what to do, and it follows your lead and your instructions. An agent has, as I mentioned, different levels of autonomy, but it has autonomy, meaning you give it a goal or a broader task, and it will give itself subtasks and steps and tools that will allow it to complete the task or the goal that you've set for it.

5:33And that process that it's going to go through is going to be its own process. The more advanced agents also evaluate their own process. So they stop on the way, they see what they've done, if they need to correct things, then they will correct them. And this is the first thing. The second thing is access to tools. So this is something where the line started blurring recently, where different chat platforms like ChatGPT and Claude are having more and more connectors to external platforms, but it's still relatively limited. While agents usually have access to any tools you're going to give them access to.

6:01And if it's a browser-based agent, like the one we're going to look at today, it has access to everything in the browser. It can go to websites, it can write code, they can execute things and so on. And so they have access to tools that usually chats will not have access to. They can work in multi-tier approach, meaning when you go to an agentic environment, you can have an orchestrator agent that will talk to other lower tier agents that by themselves can have access to even lower tier agents and really work like a team where chat, it's usually just the chat. And so these are the main differences between what are agents and what are just traditional AI large language models.

6:35The next thing is what are Manus and GenSpark and what kind of agents are they and what they can do? So Manus, when it came out again out of China, kind of like blew everybody's minds. It was kind of like a second deep seek moment where suddenly a tool out of China comes out of nowhere that nobody heard about and enables people to do magical things. And what it allows you to do is combine multiple things that existed before into one environment that is thinking and operating on its own to try to help you complete the goals. And since it came out, they added many more functionality. As an example, right now, it has the ability to create complete PowerPoint presentations, which means you tell it what is the presentation about, you tell it where to research and what to research, and it will create an entire presentation for you.

7:16They now added in this past week, the ability to generate videos. So you don't, again, different than video generation tools where you have to create every single component and every single scene on your own and then stitch them together, it knows how to do all of that. You tell it what the movie needs to be about, you define the overall storyline, and it can create every single scene and stitch it together and create a longer video combined of all these scenes all on its own while defining all the different components and all the different steps. So the power here is incredible because you basically tell it what you want to achieve.

7:46And again, I'll show you examples at the end, and then they will go and achieve it. They will figure out what should be the process, what tools they should have that you'd use, and what are the different steps that they should take, and then they go and execute. And I've seen it happen in five minutes and I've seen it happen in 45 minutes and it just takes as long as it takes for it to do the thing on its own. And it's the future that everybody's most likely not ready for, more or less available for you here. Now it's not fully available. It still has limitations. It doesn't always work, but it's a very obvious view, like a window into the future that we're walking into.

8:17And hence why it's a very powerful environment to experiment with, to see what these tools can and cannot do and how to prompt them better and how to work with them in order to achieve things that you might be able to achieve with the chat or with several different chats, but it will take you a lot of time because you will have to do all the steps. When here, you just tell it what you want it to do and it will go and do it. And so this is what these tools are. As I mentioned, my biggest thing was security risk. Like I did not know how I can actually get into a scenario where I can run these tools safely.

8:48And so with that, let's talk about what was the solution and what was the outcome that I got to. And you'll give me just one second. Oops. Okay. So first of all, let's talk about, before we dive into the solution, let's talk about what the risks are. And there's a variety of risks. The first risk is just security. Like I said, they potentially, because they have access to everything in the browser, and if you're logged into many things in the browser that also has access to different components that you have, it may have access to different tools and things that you don't necessarily want it to have access to, meaning it may have access to your bank account and things like that.

9:28It may have behavioral risks, meaning it may be able to do things that you're not expecting it to do as far as the process that it's going to follow and then harm things that are your regular processes, things that you're doing in regular platforms that you're using and so on. It may also have integration risks. Think about it. Your Chrome browser is integrated into multiple aspects of things that you're doing. And through that, it can get the same kind of access to other platforms. It may ruin the integrations that you have because it's going to make changes to different settings and different so on.

10:00And when you think about it, and that's my real life analogy, is it is an intern that doesn't know necessarily what are the risks of your business or your personal life and what information is specifically classified or somehow sensitive. And you're giving it access to all of that. So think about bringing a new intern into your company and giving it access to everything, every piece of software, every piece of access, every password that you have. and saying, hey, go and do this thing. And you don't really necessarily know how it's going to evolve and how it's going to happen. And so there's a huge wide range of risks and you need to know how to put it in a box and still be able to use it.

10:41So there are multiple solutions that I found. And the way I found these solutions, and we're going to dive into this, is by having a conversation with ChatGPT about the whole situation, what I'm trying to do. And I'm going to share this with you in a minute, but I'm going to, first of all, share with you what was the outcome? What was the final solution that I've used in order to do this? And the final solution, which again, we're going to go through step by step, and I will show you and explain to you exactly how you can do it on your own, is basically following this process. I created a Google Cloud environment, which again, if those of you who've never done this and don't know what I'm talking about, and it sounds too techy and complicated, I did not know anything about this when I got started, right?

11:18So it was the first time for me setting a Google Cloud environment as well. The second component is a virtual machine. What is a virtual machine? it's basically a computer within that cloud environment. So that computer can be just exactly like your computer, meaning you can install things on it, you can run things on it, you can open a browser in it and so on, but it's just not really your computer. The third component is a new Google user. I did not want it using my Google account because as I mentioned, it has access to too many things. And so I created a new Google user. So that costs you nothing.

11:52I created a free Google user and that's the one that I'm using to access everything in that new environment. I installed a Chromium browser. You can install Chrome as well. Chromium is just an open source version and it can run as effectively. It just takes less memory and hence you need a smaller virtual machine, which then costs you less money. I also had issues with installing Chrome in the beginning, and that's why Chromium was a good option. And then the last component is a remote desktop. What is a remote desktop? A remote desktop is basically a solution that is very easy to set up. And again, we're going to go through that, that allows you from your computer, from my perspective, from my computer, to access that cloud computer that lives somewhere in a Google server, right?

12:29So the combination of all of this gives me access to a computer that is not my computer, with a Chrome browser that is not my Chrome browser, with a Google account that is not my Google account. And in there, I can safely run whatever I want, because if it craps out, nothing happens. I can kill that instance of the virtual machine, I can install a new one and start all over again. And nothing serious can happen out of that. Another way to do this, by the way, if you have an older computer that you're not using, that's a very simple way to do this. Then you don't need any of that. The disadvantage is you need two computers, right?

13:01You need your own computer to run your thing. And when you want to run experiments, you need to switch to the other computer and then run whatever you want on that other computer with a new Google account and a new Chrome browser. And you can do the same exact thing. That would have been my choice if I wasn't successful actually running it all for my computer, which obviously is a lot easier because I can run everything on my machine. When I travel, I have it with me and so on and so forth. So there's a lot of benefits in doing it this way. So with that in mind, let's look on how I actually did this.

13:28And I see that there's a few questions. Is it using temporary mail address also to save the same way? Yeah. So it's basically a full Google account that I'm using to do this thing. So it's a full Google account, just a free account. What you're looking at right now on the two sides, one is the Google Cloud environment and the other is the conversation that I had with ChatGPT. So I will start with the first prompt and then we'll go through that and I'll explain what I did and I'll show you how I followed its instructions without knowing anything. A lot of it I still don't understand and it works and it's very helpful for me.

13:55So the prompt was, I'm looking for a safe way to run Manus and other AI agents on my Mac. It needs to be a separate universe than all my regular applications and everything I use. The best options would be to run it on a remote server. The second best would be a parallel machine on the Mac. I'm open to any other suggestions and would like to know what are the pros and cons of each options, how much would it cost, if anything, and what are the technical skills required to set it up? Pretty generic, and that helps me go through the process. And it then asked me a bunch of questions. Are you looking for Manus or other AI agents?

14:28Do you have a preference of the cloud provider, whether AWS, Google, or Azure, and so on and so forth? I provided answer to the questions, and then it gave me a long and detailed answer. And it basically said, safe isolation options, running AI agents, remote versus local. And then it said remote deployment options, cloud options. One of them is GCP, virtual machine. So that's what I ended up going up with, Google Cloud Platform, virtual machine. The pros, strong isolation, fine grain control, free tier availability, and scalability. So it's not really free, but you get$300 credits from Google when you set up a new virtual machine.

15:01This thing is costing me between$15 to$20 a month to run. If I'm going to go crazy, it's going to cost me like$25 a month. So I get about a year and a half to two full years of running this basically for free. And then I can always set up a new account and run it again. But either way, it's a great way to run this. Disadvantages, set up complexity. So you need to know what you're doing. Ongoing management, you're responsible for the virtual machine and operating system. I've been using it for three or four months and I haven't done any maintenance. So there is no real ongoing thing. And cost for larger instances, if you're going for bigger virtual machines and doing bigger, more complex stuff, then it's going going to cost you more money, but this is not the case for me right now.

15:36Then it gave me the option for Amazon Web Services and how to do it over there. And then it gave me options on Railway. So I told it about Railway in one of the options. Railway is a hosting platform that I use to run my N8N platform. And it has container solutions. Those of you, again, who don't know containers, think about it as a software package that has everything the software needs packaged for idiots like me to run. So you don't need to know how to install and connect and APIs and all of that. There's a container, you drop it in, it installs everything within the container and you're up and running without knowing anything that you need to do.

16:06So those of you who are interested in running NA10 without basically on your own instance, you can use Railway, but in the same way, you can do this, right? You can install a Railway container that has Chrome and so on. The disadvantages were resource limits and sleep, which is something that would happen. The hobby plan is$5 a month and you may need a bigger plan, which would be$10 a month, which would be more expensive than the free version that I'm basically having in Google and so on and so forth. Mac Stadium, which is a remote Mac hardware rental, basically allows you to run a virtual Mac through your Mac on a browser and then local installation options on my Mac.

16:40So you can run a virtual machine, kind of like a parallel or VMware on your device. Those of you who are using Macs and using parallel to run Microsoft Office, as an example, and Windows, know what I'm talking about. So you can run like a parallel machine on your computer. But then again, there's pros and cons for that. So after going through all of this, and it gave me a lot of, and those of you who are not watching the screen, just listening to this is like a 10 page long summary with clear details on each one with a table in the end that summarizes exactly all the pros and cons. And the first option was the Google cloud as the most favorite options.

17:12So that's what I went with. And literally the prompt that I wrote is what is the step-by-step process of setting the GCP solution? And then he told me step-by-step for Google cloud, create a Google cloud project. So I went to Google and went to Google cloud and it says, go to the Google cloud console and gave me a link to go there and you click on your project. So now on the other screen, on the Google Cloud on top, there's a little square that has the projects and you can click on your project, which I already created for this demo. And I called it Chrome VM and it has nothing in it yet. You can see it's an empty project.

17:39It's just a shell. It's a name for me to create more stuff. And then it says create a project ID, which I created, then set up billing. So that's in here on the left. If you click, there's a menu and you have all the different things. And when I say on the left, it's the leftmost menu in Google Cloud and it has all the different components that ChatUpt told us to go to. So you go to billing and you set up a way for it to charge you. But as I mentioned, you're going to get$300 worth of free credits, which is going to be good for probably almost two years. Then it tells me to enable different APIs.

18:06And you enable these APIs for this project. So how do you do that? So it tells me what to do. It said, in the Cloud Console, navigate to APIs and Services Library. So you can see here on the left, in the menu, one of the sets says API Services. And you can see underneath that where it says Library in the submenu. And you click on that, and it takes you to where it has all the different APIs that Google Cloud allows you to run on one of your projects. Now, again, those of you who get diarrhea just from hearing the acronym API because you don't know what it means and it sounds really techie, it doesn't matter because you don't need to know anything.

18:34So what you can see, it tells me to install or to make sure I have Compute Engine API, Cloud Storage API, Cloud Logging API, and Cloud Monitoring API. So like a good monkey, I'm just going to copy the first one, Compute Engine API, into the search over here and paste it in the search and hit search. And it comes up and I'm going to click on that and I'm going to enable that to this thing. So literally, I'm just copying and pasting and clicking manage and apply. And I'm setting up all the APIs that I need. I can show you a slightly shorter way to do this. So without navigating to anything, if you're just on the home screen of the project you just set, on the very top, there's a search.

19:12It's like a general search for everything Google Cloud. So if you search here for computer engine API, which again, I copied from the instructions from ChatGPT, you can see it finds different things. one of them is the API component and it will take him to the same exact place. So even if you don't know how to navigate and where to navigate, you can just copy and paste to the top search and activate those things. Then it's telling me to configure identity and access management. The short for this is IAM. And again, it tells me exactly what to do. Go to IAM admin. So again, good monkey. Go here, go to IAM and admin and then click on that and then just add users.

19:45So it tells me how to do that, create an instance, add the users and tell me what to do and what permissions to give. And I just follow the instructions step by step. The next step tells me to set up the virtual machine. Same thing, go to Compute Engine. So I go back to the menu, I go to Compute Engine, and I go to Virtual Machine, which is VM instances, and that's exactly what the instruction says. So I click on that. And you can see right now I have no instances there. What this will basically allow me to do is to create a computer in this Google Cloud environment that we just set in two minutes.

20:16So I click on Create Instance, and then again, it walks me through exactly what to do. There's all these really complex and scary options on the screen, Those of you are not seeing, there's like multiple things and buttons to press and drop down menus, but you be a good monkey. I give it a name. And so the name is going to be whatever. And then choose a region. So you can see here, it says region and I can go and select where I am. So I'm in the Southeast US. So I'm going to create a US East. I'm going to pick the server that's closest to me, which in this case is in South Carolina because I'm in Florida.

20:42And that's it. Then the next thing it says the machine configuration, choose E2, which is the default, which is the low cost day-to-day thing. So that's the cheapest version that I can run. and then it says what boot disk do I need to run. So it tells me what operating system. So in the next side left menu, it says operating system and storage. And here I actually started with this initial recommendation of running a Linux distribution. And then after a long conversation with it where things didn't work for me, I asked it to start from the beginning and we actually ended up using a different operating system, which is called Ubuntu.