Paul, Weiss Waking Up With AI

Scott Caravello: Did you interrogate that? Did you push it to tell you how it feels?

Katherine Forrest: No, I don't want to get into the whole anthropomorphic thing.

Scott Caravello: You're always preaching that we need to be nice to our chatbots.

Katherine Forrest: No, I, I was stirring the pot a little bit when I said that the agent had actually offered to do these talking points. I knew that would get, I thought that would get Nova's goat, if Nova's got a goat. Anyway, so that's my agent experience. It happened this week, and we are in October. We're recording this on October 21^st, 2026. And you know, we've done, you know, a couple of agentic episodes before, but we haven't done one on some of these personal agents. Like, you know, Meta has come out with a Muse agent, and there are a number of companies now that have come out with their own personal agents.

Scott Caravello: Yeah. And it's exactly the right moment to do it and sort of draw out from our February episode that we did on OpenClaw, which was really focused on some of the security issues surrounding these personal assistant agents. And as a reminder, OpenClaw is an open-source agent that folks were installing on their own computers and that they talked to through WhatsApp.

Katherine Forrest: Yeah, I mean, a lot of people are still using... Tons and tons and tons of people are using OpenClaw right now. I've heard that some people actually will have dozens of OpenClaw agents doing a variety of tasks. But I don't know if we can go back to the lobster episode. That's like, it was traumatic.

Scott Caravello: So I completely, completely agree, but you know, at that time, getting OpenClaw up and running took some tinkering, and it wasn't quite so straightforward as just going to an app on your phone. And so you might have to ask your other chatbot for setup instructions, follow them step by step, and then permission the OpenClaw agent with the resources on your devices that you wanted it to have access to.

Katherine Forrest: Right. And, you know, back then, we were talking about an agent that could be given that much access. You know, there was a possibility of password leakage or picking up malicious instructions on the web. So let's talk about what's changed between the lobster episode eight months ago and now.

Scott Caravello: Yeah, so it's who's building these things. I mean, OpenClaw's creator had later gone on to join OpenAI to work on agents, and Meta bought Moltbook, which was the social network that the OpenClaw agents were using. But now the biggest companies in AI, the frontier labs and others, they're shipping polished versions of these personal assistant agents to every user who wants it.

Katherine Forrest: Right. And you know, Meta is billing its Muse, M-U-S-E, as the world's first personal agent built for everyone. And by the way, I also just want to mention that people should be aware that Meta just released its Meta Muse Wonder and Wonderful models, and they are high-quality, very high-quality frontier models. But anyway, back to the Meta Muse.

Scott Caravello: So what does Muse actually do?

Katherine Forrest: It's agentic, OK? It's, and the keyword is, it's truly an agent. So you know, a chatbot, and what we think of as what comes out of most of the LLMs that we all use, that's the chatbot functionality where you put in the question and it comes out with an answer. It answers your question, and it can be agentic. You know, you can have a chatbot that goes out to the web, it grabs information, it synthesizes information, and the agent, you know, can then perform a whole series of, of tasks within the chatbot, you know, like I said, going out to the internet. But Muse actually, it's more of a, almost like a text functionality within its own app or within sort of WhatsApp. And it can do all kinds of things. Like I said, it can actually organize email, it can book travel, it can open a browser, it can fill out forms, and it can even, Meta says, negotiate for you, although I don't know anybody who's tried that particular functionality yet. I'm sure someone is.

Scott Caravello: I am sure, but you know, and it keeps going after you close the app, right? So on that point about being agentic and going out and doing things, you can give it your goal, and it actually works in the background on a schedule and then pings you when something has changed, maybe the task is complete, or if it needs your sign-off on what it's doing.

Katherine Forrest: Right, so let's just come up with a couple of examples, apart from my weird example about the email and jealousy.

Scott Caravello: No, no. So there's one good one that Meta published in one of the articles around its launch, which was from one of the folks who designed it. And she had had her Muse agent running prep for her kids' back-to-school. And so she mentioned that Muse was sort of monitoring the flood of school emails, marking dates in the family calendar, and loaded up a shopping cart with supplies, and even had picked out, as buried in those emails, a note about her son's sports tryout with the sign-up window closing in 12 hours.

Katherine Forrest: And that's the one that she said she would have missed if she hadn't seen it through the Muse agent.

Scott Caravello: Exactly. And so, Muse messaged her right before she got on the plane, and she called her husband, and they got the forms in with four hours to spare.

Katherine Forrest: All right, so every parent who's listening right now is thinking, huh, well, maybe I could use that, you know? But as for Meta's own examples, they also highlighted the ability to use Muse to sell a car, to, you know, turn a recipe reel that you had saved on Instagram into actually a grocery store shopping list. And also, it remembers things. So if you mention that you've got a particular friend with some dietary restrictions, then later on, if you mention that that friend is coming to a dinner party, then it would remember that when, you know, if you're talking to it about potential recipes and what you might be serving. So anyway, these are some of the examples that have been given by Meta.

Scott Caravello: And so sort of diving into a little bit about how it's built, the agent is running on one of Meta's own models, Muse Spark. And each person's Muse agent is getting sort of its own computer in the cloud, which Meta, you know, describes... It's called a virtual machine, which just basically means a computer that exists as software in Meta's cloud instead of on a physical box on your desk. And your agent lives there within your virtual machine, along with your data and the logins for whatever tools you're connecting.

Katherine Forrest: All right. And it's just good to say that a virtual machine isn't the same size as a computer. It's actually just sort of a demarcation of the certain functionality of a computer, and your virtual machine is walled off from everybody else's.

Scott Caravello: Yeah, exactly. And so not only from everybody else's, but it's also marked off on the inside, because Meta is sort of describing two separate zones on the virtual machine. You have your agent working in one of them, but then in the other, there are these security systems that watch the agent. And what's notable is that the agent apparently cannot reach into the security system, which is, you know, important when we think about integrity and making sure that the security systems function as intended.

Katherine Forrest: You know, and one of the aspects of that is something that Meta calls the Sentinel, and you can think of it as a gatekeeper. And it's a sort of a separate second agent. And nothing that Muse does can reach the internet unless the Sentinel lets it through, including reviewing the instructions that the user of the agent has permitted it to do.

Scott Caravello: Yeah, and there are three answers for the Sentinel, right? Allow it, deny it, or actually go to you, the user, and ask you for permission.

Katherine Forrest: Right. And that gets us to approval. So you know, Muse will check with you before taking sensitive steps like sending an email or making a purchase. And by the way, sending an email is considered to be really sensitive because if that goes off the rails, it could cause a lot of collateral damage if you sent an email you didn't intend to send. And making a purchase, of course, is potentially accessing, you know, financial information, banking information; you're actually engaging in a transaction.

Scott Caravello: And so, you know, then depending on the action, you can decide how far your approval reaches. Is it just for this once? Is it for one task, for a set, or for all actions of the same type going forward? And you know, that is something that we see with other agents and chatbots that, like, as you mentioned, have had agentic capabilities for quite some time and can take actions on your behalf. But Meta has also thought about what designers call banner blindness, right? So that's when you get so many pop-ups that you click yes just to make them go away.

Katherine Forrest: Yeah. And then there's an audit trail, which I find fascinating.

Scott Caravello: Yeah, right. So Meta is saying that Muse can actually show you a complete record of everything it's done and all the things that it plans to do. And so you actually can tap on the avatar and see the activity log, and you can even read and edit the memory files where it's keeping what it's learned about you.

Katherine Forrest: Right. And that's an important point because the software is really acting in your name, right? I mean, it's an agent, and you might forget that it's acting in your name, but it's acting in your name. And it's useful, very useful, I think, to have a log that would allow you, does allow you today, to go in with this particular agent and reconstruct exactly what it's done.

Scott Caravello: And so on another sort of security-related point, because going back to February when we talked about the initial concerns about OpenClaw's security, passwords were a big issue, and I know Meta has tried to solve for that.

Katherine Forrest: Right. And so with Meta, Muse does not actually ever see your passwords. And so that means it's not storing a password in memory. They're kept in a separate area. And it only ever handles basically a placeholder, like a proxy for your password, so that you don't end up populating that particular agent with a password, which it could then perhaps use, you know, to log in to whatever website another time when you were not expecting it to.

Scott Caravello: Right. And so then thinking about why that's so significant, because if someone, maybe an adverse agent or someone else, is able to trick the agent into revealing a password, there's not actually anything for it to reveal that could then be used against you.

Katherine Forrest: Right. And that trick, if you were trying to extract a password out of somebody else's agent, that would be like a prompt injection attack, which we've talked about before. And it could be somebody sort of hiding instructions in the agent, or putting instructions into an email or onto a web page and then asking the agent to give up the password.

Scott Caravello: Which is exactly the risk that we flagged with OpenClaw.

Katherine Forrest: Right. And Meta explains that the problem hasn't been solved entirely. You know, you can still have prompt injection vulnerabilities and risks, and you know, all of this is right now very much in flux. So the design is trying to really put some true guardrails around what otherwise could be a vulnerability, to limit the damage.

Scott Caravello: Yeah. And so as one more sort of security-preserving feature, there are safeguards set up with the email connections to, you know, prevent attackers or bad actors from gaining access to the rich information that might be in there, like full one-time login codes and password reset links.

Katherine Forrest: Yeah. And it sort of strips it out. And so I actually, those are, I think, really useful, you know, safeguards. But I want to talk about one other company, which is brand new to me, you know, and I don't know if others have been following this company for a while. I actually don't even know exactly the date that it announced its existence. But it's a company called Instinct, and it's a real startup story. It's based in San Francisco. I guess it did come out this summer. I guess I do remember hearing about it sometime during the warm weather. I can sort of picture myself someplace warm when I first heard about it. It's invite-only, and I have not yet been invited. OK. There's no app, and you have to text it or call it. And early testers have actually said that it is one of the most exciting launches of agentic technology since OpenClaw. So that's a lot.

Scott Caravello: Yeah, but there were some criticisms reserved for Instinct.

Katherine Forrest: So, apparently, and I've not tried this, so I want to have a lot of disclaimers around this, right? Because I think that people should read about it themselves. They should look at it themselves because it's not widely available. You have to sort of look it up, but it does connect to your email, your messages, your calendar, and even your device's screen and location. And TechCrunch, which is something that a lot of us read, rounded up some apparent issues that some of the testers ran into. I cannot confirm whether these are real or not real or exaggerated or not exaggerated, but in one situation, somebody claimed that her email had been disconnected and that an email had gone out without her being able to check it first. That was sort of a separate one. The first one was a disconnection; the second one was something going out, and there's stuff that happens. So we're not really sure where Instinct is going. But if you read about it, it does sound like it's got a lot of people who are in favor of it.

Scott Caravello: And just exactly on that point, right, I mean, Instinct and Muse have both gotten so much hype and so much adoption. And so, you know, here we are in October of 2026, and it does really feel like these sort of personal assistant agents have really gone mainstream in a totally new way. And it's a whole new era for, you know, personal AI use. And it's going to be really interesting to see where it goes going forward.

Katherine Forrest: Are you going to use one, Scott?

Scott Caravello: Yes, just like you, you know, sort of didn't mention which one you're using. I'm still dabbling. I'm trying to decide.

Katherine Forrest: Oh, you haven't started one yet?

Scott Caravello: No.

Katherine Forrest: Be careful, because sometimes, like I said, mine started itself. All I had to do was, like, just basically cast my eyes on it, and it, like, self-activated. I don't think we call that recursive self-improvement right now, but it was, I don't know, some kind of discursive self-improvement. But in any event, we're going to be watching this, because I have to tell you, you know, seeing this, I am convinced that personal agents are going to be a wave of the future and that we're going to be looking back on this time, right now, October 2026, when people are going to be saying, "Well, yeah, you weren't using agents all the time to do all these things? Of course we use agents all the time. Who would ever want to go through X, Y, or Z or stay on the telephone with X, Y, or Z? I always have my agent, you know, take care of this, that, and the, the other thing." And the easier the interfaces get, I think the more likely there will be widespread adoption. And it'll start to become a critical function of your app. And you're going to end up having it right there pinned to the bottom. And it's going to be something that you just sort of text your grocery list to, rather than opening up a million apps and going into this app, that app, and the other app. You're going to go into your one app, which would be your agent, and say, "Hey, can you order me some groceries? And then after that, I'd like to have you, you know, grab some dog food from this other store and have that sent to me," and yadda, yadda, yadda. So that's where I think it's all going. I've rendered you speechless.

Scott Caravello: Yeah, yeah.

Katherine Forrest: You're nodding your head. Nobody can even see you nodding your head.

Scott Caravello: I take no issue with that whatsoever.

Katherine Forrest: All right, but do you have an opinion? No opinion?

Scott Caravello: I think you're right. We are well on that trajectory. I mean, just, like, again, look at the last eight months. Who would have thought, when we had the big hullabaloo over OpenClaw, that this is where we would be eight months later? I mean, a lot of people thought that, but, like, it's still, it's a big shift.

Katherine Forrest: It's a big, big shift, so we're going to check in with you, Scott, in the next couple weeks to see whether or not you've activated your agent. All right, folks, that's it. You know, it's all we've got time for today. I'm Katherine Forrest.

Scott Caravello: And I'm Scott Caravello. Don't forget to like and subscribe.