From Autonomy to Observability: Running AI Agents Safely on Your Own Machine

Speaker 0: Hello? Right. Hey, everyone. So I'll start with the use case, what I'm gonna present about. So I am talking about 2 of my passion projects, which I'm working right now. Speaker 0: So this you see this just made. This is more of a remote cloud which we start which I started in November. And, of course, you know what happened after that open cloud was dropped. So and this again happened. Was that your open cloud? Speaker 0: Okay. Okay. So and that's how I started. So what I want to showcase here is that this is where I started with, but this is not what landed on. What I landed on is, this. Speaker 0: Let me mirror this. This might be much better. So what I am working right now and what I've ended up is, which is more of the observability here. So what I wanted to showcase here is, like, if you think, and this technically applies to all the remote agent you have been working on that, hey. Whenever there's a agent and, working on your machine and you're working on autonomously on on that, like, how do you control it, or how do you know what's actually it's doing? Speaker 0: And thank you for the survey before that. I think a lot of the questions on the survey are who is using local agents. We see a lot of the hands up. Who is using or leaving those for more than 30 hours or 30 Winsnes? Sorry. Speaker 0: And we have a a lot of hands up there on that. But the question there is now that, hey. Whenever I am maybe running an agent and it is trying to interact with your system and it is doing something, and this is a very simple question about your battery status, status, which I've asked, and it should be able to do that. But what goes on in the background? Like, how do I keep a track of what's happening? Speaker 0: And that's where, this, a project which I'm working on right now is, like, in terms of the observability comes in the picture. So you can imagine all these agents are running on my box right now. There are bunch of cloud code. There's OpenAI running. There's GitHub Copilot. Speaker 0: But what I want to make sure that I can go somewhere and see, hey, what all is what all they are doing. Is there any network, you know, issues which I should be aware of? And if you see, if I just say hello, here, it tells you that, hey. There is an ingress happening on this machine right now for this plot code, and then, you know, you can also see, like, when it stops. So those aspects which you can keep a track of, along with that also tells you, like, how you've configured your agents locally. Speaker 0: Do they have access to something which they should not have? Do they have 700 or, you know, 7 x x in, permission on your folders which you didn't mean to? Because, you should like, I I think we all know that OpenCloud is good, for example, but the defaults are not that great. You have to make sure they're configured in a very secure way to even proceed to like, at at least that's my personal opinion that I love OpenCloud from that perspective and do things, but I'm not yet confident that it'll do it in a better way. And this is, giving an example of eventually 1 minute. Speaker 0: 1 minute. 1 more minute? Okay. And this is where I'll just quickly show, maybe, you know, you can go through this, and this is how I'm defining forensics here as well, that the underlying story is, like, there are tools which can give you traces. You have h stop, which can tell you what for a process is running. Speaker 0: But what I'm I'm trying to do here is to have that single pane of, you know, visibility where you can go and see venue interacted with your agent, what are they starting to do right from the reasoning, what process did it touch on your local box. So go through that forensic, and eventually, how you can share this forensics between each of the agents so that they can do a better job. So that's how I'll stop. This is a open source project, so collaborate. Like, looking for contributors, feel free to go on the website and start, coding it. Speaker 0: Thank you so much. Speaker 1: Any question? Yes, sir. So is is this just for OpenClaw, or is this generically something that people can put as, like, an agent on computers to monitor any egress? Speaker 0: Yes. Good question. So this is generally for all the agents. So, today, I'm supporting I think this project is supporting 14 agents out of the box. The the main point which, is helpful here is that this is a drop, tooling. Speaker 0: So you just drop this on your box. It will start picking up all the signals signals, which, you know, multiple agents supports, and that's how we've configured Speaker 1: it today. Okay. My question is, what's your worst horror story? What's the worst thing in the past? You said, oh my gosh. Speaker 1: It's doing what? Speaker 0: Good point. I think that's how it started. The worst which I felt, and there would be really, you know, was the incidence of that is, like, there was an agent which I saw was running on my machine for the last 12 days, which I didn't even know about. Okay. There is an agent there. Speaker 0: And when I actually brought this, I'm like, okay. That says, uptime 12 days, some hours, and some minutes. And that's where it kind of, made me to think more deeper than, hey. What was it doing? I didn't have logs. Speaker 0: And that's where I go and introduce forensic to see actually what it is doing. So that's the worst for me, but, yeah, it could be anything. Like, it Speaker 1: has taken some action Speaker 0: that you don't know about as well. Speaker 1: That's hilarious. You're on your side. It's on your box. I did see that you had some, unprotected credentials Speaker 0: that were stored in in plain text when Speaker 1: you were going through. So, yeah, better better check on those. Speaker 0: That's what this is for. Right? Speaker 1: For you to know what Speaker 0: you're doing before John looks at your machine. Speaker 1: Yes. Yes. Thank you, folks. Alright. Speaker 0: Yeah. Have a great round of Speaker 1: applause. Yeah.