The 5-Step Framework for AI Agents That Improve While You Sleep | E2269

All right, everybody. Welcome back to this week in startups. It's Twist for March 30th. Happy birthday to my twins. And it is. They're 10 years old. We had their party this weekend. And I'm in Brooklyn. I'm in Brooklyn. I came to see clearly.

0:00

Yeah.

0:14

As you can see from the Brooklyn Bridge. My dad's been a little bit sick, so came in to spend some time with him. See the fam, see John the beer.

0:14

That is not a live shot of Brooklyn behind you. I. I call, I call shenanigans on this.

0:21

Well, yes.

0:27

The wrong time of day, sir. It's the wrong time.

0:27

I mean, it's a time zone. I think Texas is what, eight hours behind?

0:29

It's like Tokyo. Yeah. It's like eight hours. Yeah.

0:33

This week in Startups is brought to you by eru. Unify your identity, endpoint security and compliance

0:39

into a single platform.

0:45

Twist. Listeners get 20% off when they book a demo@erru.com Twist that's iru.com Twist LinkedIn jobs hire right the first time. Post your first job and get $100

0:46

off towards your job.

0:59

Post@LinkedIn.com Twist and Quo founders move fast. Their phone systems should too.

1:00

Quo formerly OpenPhone gives you a clean,

1:08

modern way to handle every customer call, text and thread in one place. Try it free@quo.com Twist that's Q-U-O.com Twist all right, we got a big show today, but before we get started, this is what we do. We get ready for the show and I applaud Plod by pressing my pen like we are in Star Trek. And it really is kind of like Star Trek. One of the things I wanted to do was I wanted to interview my dad. I'm trying to convince my dad to let me interview him so I could write like a short story about him or something down the road. And, you know, he's an old school guy. So I was like, I don't know. My story has been told.

1:10

I don't know.

1:53

But I was like, okay, well, I got this plot pen. I just want to talk to you. Like, maybe we could carve out an hour or two to just talk about your childhood or whatever. I don't know.

1:54

You got to capture those memories. Once people are gone, those stories are gone. I feel this with my grandparents all the time, that I wish I'd gotten more of their old stories because now they're gone. I'll never be able to ask them about the old days.

2:03

And I bet you somewhere in the plodding templates is archivist or, you know, family archivist or storytelling templates. If there isn't, we could write one. And that's the beauty of Plaud. And, you know, they're a partner of ours and they're sponsoring today's live stream. Go check out Plaud. There's a pin, there's a. A mag safe adapter that goes on the back of your iPhone or if you got any other MagSafe.

2:15

Yeah.

2:38

It is extraordinary. It's changed my life. It's privacy first. So you see the red lights on, and it's changed along his life. Even Alex loves it. Now he's getting into it.

2:39

It's great because it's, it's, it's. It's not just recording, but it, it understands the context. I mean, I think that would be great.

2:48

Why?

2:53

It'd be good for what you're talking about. It would not only capture your dad's story, but his, his tone, his personality, his voice. It's really, it's, you know, it's good for the contextual understanding of what's being said.

2:54

And, you know, I think these things are going to keep getting better and better. I have a feature item I want for plot. I want Plaud because it is independent of your device. If you have your phone in the charger and you walk to the kitchen and you want to, you know, do like a quick session or something, you know, with a colleague in the, in the corporate kitchen, you can just press it and record it. Yeah, I wanted to have, I wanted to pair with its own proprietary earpiece to give, like, feedback. Or I could ask it questions back and forth and then eventually have a local language model on it and have it connect to WI fi. I don't know if you knew this, but if you put it in the cradle, this is like a power user feature. If you put it in the cradle, it will sync over WI FI for you.

3:05

Yes, I did know that. I did know that.

3:46

You did know that. Okay. It's like a little secret feature because I was like, you know, this random thing. Yeah. I'd like this thing to automatically sync. And like, the perfect time to do that. When I put it in the cradle. What's the URL? Help me out here for a special offer. That's.

3:47

Yeah. So you want to check.

4:00

We can talk until you give us the privacy.

4:01

I understand now. So you want to, you want to check it out. Applaud that. AI/twist. That's plaud dot. AI/twist. You want to use the code twist for 10% off of your plot pin.

4:03

All right, so it is AO after

4:15

open, we stop keeping. Got to be. We got to be close to the 60s. Yeah, we're in the.

4:18

So we'll just go with 60 for today. Open Claw is a transformational product that allows you to have an agent and that agent, you can give it access to everything. You can give it its own computer, you can give it its own Gmail address notion. And a lot of folks are now amping up their products and services. I noticed Claude, I noticed Claude added specifically skills. So you can add a skill to Claude Cowork now. So this is such an impressive product and an open source project that our friend Dave Morin has been pursuing and I think it's really important for us to keep the open source train moving. So I'm keep talking about openclaw, but I'm actually delighted to see that the innovations in there, specifically skills, have made their way into other products that we love. Perplexity Computer I think has added skills. And I just noticed Claude, if you hit skill, it will let you put a skill in and then you can copy and paste the skills from one to another. One of the skills you and I created that was super helpful to us and is obviously doing guest research. Great, right?

4:23

Yes.

5:28

Another skill is, hey, what are the best practices around thumbnails and subject lines? Like social media? Best practices for YouTube, whatever. So now we have that skill. I have a skill for tell me what the person's done, like a check in report, like a management tool cir. So I could say check in report for Lon3 and it'll tell me the last three days of what Lon worked on. In other words, the weekend. And I'll say he takes a full Sunday off, maybe checks in once in a while, but he tries to keep that Sunday nice.

5:28

And I try to keep my Sunday. I work a lot on Saturdays.

5:55

I feel you do put in some time on Saturdays, which I appreciate. Yeah, but it will tell me all the Gmails Notion Zooms meetings. Anyway, we thought we'd step back a step today, have a guest on who's extremely good at OpenClaw, just go through their top five open claw tips. So it's not just us repeating our tips over and over again. So let's introduce our guest and get to work.

5:58

Yeah, we're very pleased to welcome. He's a senior AI product manager at Google, but he came to our attention because he's been posting incredible open clutch tips on X. A lot of great X articles. Including one that I bookmarked called How I built an autonomous AI agent team that runs 24 7. Please welcome Shubham Sabhu to the show. Shubham, thanks for being here.

6:20

Thank you, Jason. Thank you, Lon for having me.

6:40

Yeah.

6:43

Yes. Super excited to talk about OpenClaw. I've been running OpenClaw for last two months. Today it is a team of six agents running on my Mac Mini. They pretty much automate everything that I do outside Google. So quick context for the viewers. And for you, Jason and Lon, outside Google, I started an awesome LLM Apps repo which was an open source LLM Apps plus agent templates for users to build on top of. So they can just use those templates, build on top of that. Repo crossed 100,000 stars and it was such a pain to manage a single person site shop. Apart from that, I have newsletters and whatever I learned with these two goes on my social media. So that would take hours and hours of my time outside Google. So I was pretty much working like 20 hours or 18 to 20 hours before Google. Then I came across this thing called OpenCloud. I did not use it for 15, 20 days since its launch. And then I see it popping up on Twitter X and LinkedIn and Reddit everywhere. I was like, I have to try this thing. Like, what, what is this thing? I install it. I. And first of all, for viewers listening, you don't really need a Mac mini to install OpenCloud. But I am super concerned about my security. I keep these things seriously. So I did not want it to install it on my personal machine as like $500 is worth an investment. I got a Mac Mini.

6:43

Can I, can I jump in here and ask? Because we have our Open Claw agents running in an AWS EC2 instance. And I do, I have to say, I feel like it's holding me back. I feel like I encounter a lot of like, oh, I can't go to that website. Oh, I'm getting blocked. I can't look at that video. Oh, my email doesn't work right because it's because of the limitations of that. So how, how big of a whole thing holding me back is that and is it really like preferable to have it in its own Mac Mini or some other solution?

8:09

So it does not really need to be its own Mac Mini. But what I recommend, what I've seen after talking to a lot of individuals is having OpenCloud running on a clean slate or a clean machine really helps because what you're trying to do here is trying to give agent its own identity, personality. So you really want that agent to have its home as well. Just like Lon, you could live in a rented apartment, but you would really prefer to have your own home, right? Autonomy, privacy, you can do whatever you want. So when you give OpenClaw a clean machine, be it a Mac Mini or a local or your old computer, you can just leave it autonomously running and doing like changing files and using websites, attaching it to Chrome, using browser. So that kind of flexibility comes with a machine which you typically won't get in sandbox environment. So they have its limitations. But people starting out, you can just start in $5 or $10 subscription, just run it in sandbox environment so you really don't need to buy a ma. That's the first step. But once you get claw pile then you would really want to have a Mac Mini because that's the cleanest user experience you would have. So I've been doing that and When I installed OpenClaw I just realized it is super simple to install. It is just like installing OpenClaw, selecting models, giving them tools, going through certain steps and by the end of it you just have your first agent running. It comes with a UI along with the terminal. So you can just go in the browser, open the dashboard and you can see your first agent running. And that's how it started.

8:42

It's springtime, folks. And that means it's time for spring cleaning. But not just for your house, but all those messy parts of your business. And that's why today's episode is brought to you by Quo. The smartest way to run your business, communications. Even the most attentive and professional team members, well, they're just human. And sometimes we miss a phone call. That's why more than 90,000 businesses rely on quo to stay connected and consistently reach you. That's because your entire team uses the same number. There's no more slip ups or people using their personal phone numbers. And then they leave the company and you lose the entire history of the customer and your partner. Plus Quo has AI that automatically logs, generates summaries and recommends next steps for your calls. So make this the season when no opportunity and no customer slips away. Try quo for free. Plus get 20% off your first six months when you go to quo.com twist that's Q-U-O.com twist quo no missed calls and no missed customers.

10:14

And when I was starting out, I see people naming their openclaw agents randomly like Sam or Holly or Molly. And all Those things. I was like, it is very difficult for me. Like, okay, it's good for one agent. I could remember it, but I was like, maybe let me just name my agents after my favorite TV character. So I started naming them after my favorite characters from Friends and the Office. And after a while, I realized while I was doing it unknowingly, it really makes a very good mental model. So I started with my first by having Personas.

11:15

Personas create a mental model. So this person's a producer, this person's a fact checker, this person's an engineer, this person's a researcher, this person's an accountant. That does help humans figure this out. But then there's the Ultron model, which we've talked about here is like having one have all the skills. And that's an ongoing debate. And I think we've all come to the conclusion, Lon, that it's fine to start with multiple agents and coordinate them. It's also fine to try to have one super agent. You'll hit roadblocks in terms of how often you're using them. I think that becomes the key blocker. Yes.

11:49

Yeah, that's 100% right, Jason. So I never recommend people. So people see all these flashy screenshots on X where people share. Okay, I've been running six agent team or eight agent team, or I have, like 12 agents running in the background. I never recommend people to start from there. Start with one agent and make your way up by talking to your agent. That's the best thing that OpenClaw has done. So you really don't need to be a technical expert or you don't need to be a coding ninja to run these agents, deploy them, or make multiple agents out of it. So I started with this main agent. I named her after Monica because Monica is known to run a tight ship. She handles the operation, all those things very well. And I started talking to Monica. Monica. So here's tip one. So when you are starting with.

12:27

So this is tip number one of your five tips.

13:12

Yes. When people start with Open Cloud, when people start with their first agent, because of all the hype that's created and that brings people into this, people think, okay, installing is the only thing I need. And then I just start chatting with my agent. Agent can do whatever I want. That's not how it works. So I've seen people operating in, like, two extreme. Some people just, like, directly get in without giving any context. So you really need to tell your agent about yourself. What do you want it to do? Think of it like onboarding a real employee or an intern. So you do not onboard them by not telling them anything. You don't onboard them by telling them nothing. Also you don't just like bombard them with all the company information on day one. So you don't just like dump files and files of contact. Don't just give them like here are like 10 folders, figure this out yourself. So people are operating in two extremes and both of them end up in a disappointed shape or form. And then they're like, oh, this open cloth thing, it was so hyped up and it is not working. It's like either you don't give it any context or you're just dumping it with a lot and lot of context. Because at the end of it it's the model which has limitations. If you dump it with a lot of context, it will do a context blow. If you don't tell them anything, it will just give you a generalist result. So you have to give so example

13:15

of this, example of this good context to give it and then overwhelming it.

14:33

Good context will be just like how you onboard intern. You would tell them who you are, what the company is, what the company does. You give, point them towards the right links. You tell them specifically, this is your job, this is what I want you to do. Just like Jason, you are mentoring, right, at the startup, like onboarding an employee. Yeah, yeah. You were very specific. Lon, I want you to do this and Jacob, I want you to do that. So you need to be very specific. You can't just say Jason. Okay, Lon, Jason, figure out what went wrong and just do everything yourself. You are very specific in your direction so you need to have that kind of direction and that's the right context.

14:38

Have you done that? By the way, Lon, on yours, on Gaff, have you done that?

15:12

Yes. In fact, we had another guest on Jordy Coltman. You remember this was during the very early days of openclaw and his recommendation was you have your agent interview you and then base its sole MD file on that interview. So I actually had a whole like 20 minute conversation with Gaff where I explained the show and how I'm going to use him and what kind of other tools I'm using and that really informed our whole relationship moving forward to there. So yeah, I think this is super important because you don't. Without that context, your agent doesn't know what you're looking for.

15:15

Exactly.

15:49

I actually just saw a skill that I will find and post here where actually you Set up your agent so that it always interviews you before a new task. Because the theory is when you just tell an agent to do something, it's going to get like 70, 80% of the way that you of where you wanted it to go. But it won't get that 100% because it's guessing on a few things. And so what the interview idea is is if your agent chats with you beforehand, asks all of its questions, it doesn't have to guess. And you'll get a lot closer to 100% on the first try, which is an interesting idea.

15:50

And should all that become part of their soul or part of their memory in your mind?

16:26

That's a very good question, Jason. And here's the interesting answer. So I did not tell Monica where to put all the information that I'm sharing with her. I shared what I want her to do. I shared about me. And she figured out like what to put in her soul. She figured out that she should create a user MD to put the information about me. She figured out what will go in her identity file. So this is taking us as close to managing a human or as close as mimicking a human experience. Which was the, which was super interesting and insightful thing. When I used openclaw and by default the way with the new updates coming in, by default it is going in that direction on which you're talking about. So when you throw in something at openclaw it asks you a flow follows some follow up question and then it goes from there and then it automatically figures out where this information goes in what file. We'll come to memory in a bit, Jason, but this is when you start like this is how you set up your first agent. Giving the right context. Agent figures out itself and all you're doing is talking to your agent. And that leads us into tip 2 or the productivity tip too. Because all these tips will build on themselves following a specific user journey. When you start, like how you start and how you evolve your agents into something towards the end of it will get to an agent team which self evolve, self learn without keeping you or without you being the barrier to that. You don't have to like Jason, just like you came in and you told Lon and Jacob, okay, this needs to fix. You don't need to tell them, what if it gets automatic? What if they could just keep doing it themselves? Similarly, what if agents can just figure out, okay, these things are breaking, go back and just, just do a recap of what went wrong and figure out a fix update Their instructions based on that. So we'll go from here to there.

16:31

So the top five so far we got start with just one agent and onboard it like a new hire. And then tip number two, you just got to talk to your agent. Don't worry. Sometimes I will say my agent will try to be open up your terminal and do this and that. And I always shut him down. No, Gaff, you got to do it. Learn how to do this yourself. Now we're, we're moving on what is topic number three? We're going to talk about cron schedules.

18:28

We talked about setting up your first agent, we talked about the value of talking to your agents. But still you are still in proactive role. Like you are prompting agents to do something. You are still asking agents to do something. But the real value comes from agents being autonomous. Can agents do something without you even having to ask them to do something first? So can they do something while you sleep? Can they do something without you prompting them? That's where heartbeat or cron schedules come in. That's one of the best thing that OpenClaw has introduced. Now I can just put my agents on a cron schedule. So I have this agent called Twite twice from office. He takes care of all the research because previously what I had to do is I have to be super active on X and Reddit and all those places and looking at the blogs to figure out out what's happening in the AI space. And since with the speed at which it's moving, it occupies a lot of time, you need to be like online most, most of the time hiring can

18:52

be its own full time job. And hey, guess what, I already have a full time job. I make podcasts and I invest. But when you're running a small company, we both know every hire matters. You don't want to waste any of the seats you have at your company. And the best partner you can have is LinkedIn Hiring Pro. Why? There's a billion people using LinkedIn. All the great talent are there. If you're proud of your work, you're. You build a LinkedIn page and you update it. LinkedIn Hiring Pro is gonna streamline and simplify the entire process for you. Nearly 60% of companies using LinkedIn Hiring Pro, you're gonna get an incredible candidate to interview in the first week. And you know we're looking for a new producer for the pod. We did shout outs here on the show. We posted it on my social media. We asked friends, you know where we found our Next Great hire. LinkedIn and it was competitive. We had like three or four really good choices. So hire right the first time, post your first job and get a hundred dollars off towards your post@LinkedIn.com hiringpro that's LinkedIn.com hiringpro offer. Terms and conditions apply. So this is a great example. We should if the producers can pull up the cron job schedule and I'll run through it here. So you have DWight scans at 8am 15 different sources, things like X Hacker News, Reddit and it writes an intel report and sends the summary to Telegram. Then you have Kelly at 9am read Dwight's intel and draft X posts for for you. And then they send those to Telegram for you to review. Then at 10am, Ross reviews the open PRs on awesome LLM apps. That's your open source app that has 104 stars and flags. Which ones need attention? I guess they prioritize it. Dwight comes back in at 4pm and does an afternoon sweep, catches anything new since the morning. And then Kelly and Rachel Draft X and LinkedIn post from the refreshed Intel. And then at 6pm you draft the daily newsletter from Dwight's intel and send the draft to yourself for a final review. That's it? Yeah, that's the process here. And so this would be like if you had 1, 2, 3, 4 different employees, except these are agents. Now these agents are each running in a virtual machine, their own open claw or how do you just technically architect that?

19:52

So when I had my first agent, Monica, I told her, monica, I need to keep doing this kind of research and we may need like someone who could just like scrape these things for me, who we can give these tools and can automatically, autonomously throughout the day. Monica helped me through the process. Monica's like, we can create a new agent. Then I talked to again, second tip, going back to the second tab, always talking to your agents. I gave her all the context. She gave me few options. Then we end up, both of us decided we need to have another agent. She walked me through the process.

22:09

That means you need to have a second agent. And this is where that second point is. Instead of trying to figure it out for yourself, just ask your agent because they know how to find the answers. I had this happen. I was like, hey, what's the best way for me to download videos, rip a video from Instagram or Twitter and get it into my and then store it somewhere. And we figured out or it figured out like three or four different ways. One of them was on my Mac Mini to go to like these Russian websites that rip YouTube videos and constantly get blocked. Then it was like, oh, that's not working. It's not consistent. We should probably just run this GitHub repo that was based upon what those websites are because it's open source software. I was like, yeah, that sounds great. Now I just go to like three of my replicants, Lon's, myself and another one, and I just say, hey, can you get me this video? It downloads, it puts it back into the slack room, and then for larger ones, it puts it into a G drive. But I figured that out not by going and doing the research myself, just by asking the replicant. Talking to your agents is all you need. Point number two, three, put your agents on a cron schedule and then you can really see the power of them. Number four, add cross agent memory. When you start scaling, this is something. It's interesting that you bring this up that we. Oliver, who is the producer of this Week in AI and does the demos, they're really talented. One of my new all stars and he just built a little skill to do this. Explain why this is important. Why is it important? Point number four, add cross agent memory. When you start scaling, when you have

22:42

these multiple agents, we got to shape and form where now we have an agent team and there's a chief of staff and agent hierarchy, Monica, managing these agents. So when you have a lot of these agents going on and one agent feeding to another agent, what happens normally is. So generally before I get into this, I'll just say, like, openclaw has a really good default memory layer that coordinates through file, that works through file. And a simple plain English example of that is Dwight does all the research, he writes whatever, he finds his signals in an intel file. Kelly and Rachel reads that intel file. Pam reads that intel file, creates the newsletter out of that. So that's how the coordination work. That's also your memory. So they write it in files and through files. Each agent discovers those files are memory MD file, there's daily log memory files. That's how it works. I realize what happens. Many times my agent would just forget. Forget. Like Dwight would just forget the tools that he would have access to.

24:19

This happens to me all the time, where I just have to remind be like, no, Gaff, you have that already. We already set it up. Go look. And then it'll be like, you're right. Like, I find I run into this all the time.

25:16

This is the difference between a professional company like Anthropic making Claude or Perplexity making Computer Claude. Cowork and perplexity if they did. If it didn't work and it wasn't reliable, people who are paying 20 bucks or 200 bucks a month long would then email the company and say, hey, this is not right. It's giving wrong answers, et cetera. So they might be slower, but they're going to be more, less brittle. Open source projects tend to be a little less stable or reliable in this early stages of them. And there's not like you can call customer service and be like, I paid $400 for this for the last four months and it's not working. I want a refund. Right. There's no refund for OpenClaw. So I think that's an important kind of thing here. And it's important for the people at the OpenClaw foundation to know reliability is something like adding features. Yeah, great. It's the most fun in an open source project. The features is what everybody wants to work on. Reliability is what people don't want to work on. That's chores and grunt work. It's got to be more reliable. With the memory, how do you store the memory between multiple agents? Do you give it a shared drive with the memory file? And then how do you share skills? Do you sync them across the skill files or you point them all to the same hard drive or directory somewhere in the cloud? With the skills, what's the practice you use? Because I've heard many different ways to do this.

25:26

Yeah. Before getting to a good reliable working solution, I did a lot of hacks here and there to figure out like. So make sure things are. Let's say if my agents were forgetting tools, there was a Tools MD file that were created which would list all the tools so they won't forget. And in the SOL MD we'll have add a line in dwightsoul MD always go and look at the Tools MD file. So these were like hacks that we were doing which kind of again worked and generally works for 80% of the use cases. But then for actual memory. And this is the problem that does not that did not just came with OpenCloud. It always existed since we had agents. And you would see some of these companies building solutions for agent memory. So Google has this very nice solution called Vertex AI memory bank. There is a company called MEM0, there's a company called Cogni. All these operate at the agent memory layer. So how it works is so Google we created a plugin for Vertex AI Memory Bank. It is a solution which lets your agent automatically recall and Capture memory at runtime. So when I'm having a conversation the agent automatically figure out that memory layer automatically figured out what's worth saving it auto captures that and when I'm having when you're in the conversation and it needs to be pulled out that's called auto recall. It automatically figures out this need to be recalled. So agents always have this memory and why this makes sense in across agent team is let's say if I give feedback I was talking to Monica and I tell her Monica I don't like a lot of the use of emdishes all the time so don't use emdes or use it like very very carefully. What happens is I just told this to Monica but because this is a shared memory layer, next time when I open Kelly's draft or Rachel's draft I don't see those amdations. So that's the magic of shared memory layer. That's the magic of an intelligent dynamic memory layer which of course is an addition on top of your openclaw layer which you will be paying for. So this does not come for free but there are elegant solutions already designed there. All you need is a plugin to OpenClaw. Again I am not an engineer, I did not know much about the engineering side of memory bank but then I worked with my agent to build a plugin open sourced it for others to use. So now you just pay 8 to $10 more on top of what you're paying and you have this elegant memory layer. And even before this I tried multiple things like I was like okay can I use a local memory layer where I have an embedding model running in my talker and that could work but after all those hacks I was like let me build this plugin, let me use that. Use the elegant solution that Google already has.

26:54

Start with one agent onboard it like a new hire. That's number one Lon, talking to your agent is all you need to do. That's number two. Number three, put your agents on a cron schedule. We know about that. Number four cross agent memory when you start scaling number five let your agents get better on their own. This I've never heard of. Explain to us what your thought here is on making them get better on their own. It might seem like your startup can save some money by using a cheap tool for your device management and security, but you'll eventually end up paying out the difference and more in time and frustration. Take the guesswork out and trust our friends at ERU Endpoint Management. ERU is Going to automate your device management end to end so you don't have to worry about onboarding endless app updates and enforcing all of your company's various policies across your entire team's Mac, Windows and Android devices. This is going to cut down all the time it takes you to deploy new devices from hours to just minutes. And that means less time troubleshooting, fewer tickets, and better security for your company's valuable data. Everyone at launch has a phone, a laptop, I'm even buying. And these kids plod devices now because they're so good. You think I have time to worry about what apps they're installing on these things and what websites they're visiting and all the error messages they're going to get. I got to make podcasts all day long and make angel investments without eruse system. My team would be wasting hours policing everybody's hard drives instead of doing valuable work that actually moves the ball forward. Twist Listeners get 20% off when they book a demo@iru.com Twist that's eru.com Twist

29:31

this is the most fun one and this is the one that I recently realized. So what was happening is I would found myself giving a lot of feedback to my agent. Fix this or fix that or this is not working, that is not working, that format is not right. Or Ross, you reviewed the PR or you brought me these PRs which could be rejected at the first glance. Why are you asking me to review it? Why are you wasting your energy in reviewing it? So I was like just like how humans work and taking lessons from the management skills. Because since this is a company, this is an org which is managed by Monica and these agents are employees. That's the mental model that we are going with. All the organizations have this self review process, managerial review process. Why can't agent go back and recap what they did in a week and review what went wrong and fix themselves and why can't that become a cron? So that was the seed for self improving agents. Agents can get better on their own. Now each of my agent has a self review weekly self review process where they go and look at everything that they did in the week. Kelly would go in and open X. All of these agents have their separate accounts so nothing. They don't touch any of my personal stuff. But because my tweets are personal they can go and see what are the tweets that Kelly post that Kelly suggested that I posted on X. What was the hit rate? What did well, what did not do well? Same Ross can go in and see the PRs that he suggested, what actions I should take and what I did. What was the difference between that? So Ross can go and see. Okay, this is where I went wrong. This is what Shubham did and automatically fixes its instruction. And on top of all of these, just like we have managerial review, there is a bi weekly review by Monica for all these agents where she goes in, looks at what these agents did in last two weeks and grades them like a B B C and reports send a report back to me so I could see like as a CEO I could see how my agents are doing, what is going well, what's not going well and manage them from a managerial position rather than getting more hands on. So we went from installing agents to managing agents in these five steps. That's the entire thing.

31:05

All right, great job. We're going to continue to follow you. Where can people find more? Where can they follow your GitHub? Where can they find you on X?

33:25

My X account is like Shubham Sabu. You can just like search find me on X. Awesome. LLM apps is the GitHub repo S H u P H a m S a B o o S a b

33:33

o o Is there a dash between the two words?

33:45

Or underscore or just straight up underscore?

33:48

Underscore.

33:51

Should we talk about journalism and the New York Times or should we go on to our next demo? What do you think?

33:52

If you want to do it real quick, I can basically give you my contact. You tee it up for everybody. You teed up. It sounds like I'm talking about my. I didn't know you're going to put this on the docket. I don't usually like to talk about myself, but this one went quite crazy.

33:57

This one went very big. 38.6 million views on this tweet that I'm about to pull up for March 22nd. Founders, take my advice. Do not talk to the press, go direct and do long form podcasts. Wired and the New York Times are as biased as Fox News and Ms. Now, excuse me, I was about to say ridiculous non non name for that cable network. Ms. Now this is a function of their need to pander to one side to survive be it through $3 a month subs or rage baiting ad based stories. So some of founders, if you talk to the New York Times are wired. They're gonna trash you. They're gonna misrepresent you in order to get more subscribers. So the context around this was Wired has a new editor, the New York Times Did a profile about Wired's new editor and you were saying, people don't care about this ultra woke biased version of Wired. Fewer people care about the New York Times take on Wired. Founders should just not engage with the media and go direct. So this set off, you know, on both sides. You had a lot of defenders, I would say Elon, notably Joe Lonsdale, Mike Solana, a lot of other tech leaders and CEOs said they agree there's no need to go talk to the press anymore. But then a whole lot of writers, podcasters, journalists took strong issue. I would say Jordan Crula, she's a writer and a podcaster, she had the tweet. I think that sums up the opposition the best. Wired is on an absolute tear. They're adding tons of subs, they're breaking headline news more relevant to the nation's conversation than it's been in years. If you don't like how Wired writes about your company, maybe your company should be less crummy. Although she didn't say crummy. She used, she used a bad word.

34:10

She used the S word.

35:46

She used the, she used the S word. So, I mean, I guess I would throw it to you, you know, do you, do you stand by this idea? What, why do you think, you know, is Wired having this breakout moment? Is the Times having this breakout moment? And, and why do you think there's no need for founders to talk to what we would think of as, you know, the mainstream?

35:47

Yeah, I, I'm not dunking on the press for the sake of dunking on the press. I'm giving specific advice to founders in the context of 2026. Wired magazine and the New York Times made a choice at some point, I think during the Trump first presidency, 45th presidency, to, you know, go more world to be the resistance, which is totally fine. They're allowed to make that editorial choice. MSNBC and Fox News proved horseshoe in the horseshoe theory. If you make a horseshoe, the tip of the HorseShoe are the two radical extremes. And the big wide part of the U is all of us in the middle.

36:05

Yeah, it's the center.

36:47

But those two tips, when they're enraged, they click. And when they're enraged and they're pander to, they subscribe. You can see this on the right acutely with Megyn Kelly and Tucker and then I guess to the right of them there's you know, Fuentes and whatever. Sure, you, I mean you, you get, I mean like really far to the right. Like Nuremberg, far to the right. By the way, great film I like to turn. I enjoyed Nuremberg. I am giving advice to founders. Why am I giving this advice? I would give a similar piece of advice. If you were gonna go on Fox News or something and you were, I don't know, you were at an elite institution like Harvard, what they consider elite institutions and they consider you the problem. If you go on Fox News, they're just gonna dunk on you. It's not gonna turn out well for you. It's gonna be a setup, basically.

36:49

I mean, isn't the press supposed to be a little adversary? Like, isn't that kind of baked in? If you're covering somebod, you shouldn't be just like totally in there.

37:40

Adversarial or advocate, two different words. Adversarial means you ask hard questions, you ask hard follow ups, and you talk to as many people as possible. Advocacy means you pick a side and you write stories that appeal to that side and that build that side's arguments. So if you were writing a story about the border, you know, we all know what they would say about Biden and Kamala. We all know what they would say about Trump and they would say about ice, etc. You know, they frame either way. The problem in media today is if you go right down the middle, you get no ratings because that horseshoe, you don't make any money and you lose audience. That's why people picked a side. Founders get caught up in this. If you go through like dozens and dozens of stories, whether it's about me or anybody else in tech who's high profile, and you look at a New York Times or a Wired story in this wired 3.0 era, it's going to be basically holding truth to power. It's going to be a left, you know, trash the company with a bunch of anonymous sources. And that's the other problem here is the journalistic technique of using anonymous sources has been abused in order to do advocacy. I know this because I worked in journalism, like for half my career. When people used to bring us a source, we'd say, well, what's their ax to grind? Now they're like, let me find the sources with the axe to grind. That's literally what they do. They're like, where can I find people who used to work at Anduril, used to work for Jason, used to work for this company? Where can I find people that he fired? Where can I find people who are competitors, who, you know, he beat? So my advice is strictly for founders. It's not worth your time, and it is worth your time to build a direct. Or it's. Why do I say that? The average New York Times story gets like 100, 200,000 views. You can very easily. As you said, I got 38 million views on that one.

37:48

Yeah.

39:32

You know, and I typically get 50,000 views on, you know, 50 to 500,000 views. Now. That's a million followers. And it took me a while to build it, but across my podcast, millions. That's the better strategy for founders, is just be a vocal founder, be a vocal venture capitalist, which everybody has sort of copied me on that way. Right. I was one of the first in the VC community to do it.

39:32

Yeah.

39:51

With this very podcast 15 years ago. So that I'm giving advice that's in the best interest of it now. What shocked me was the New York Times should just shut up and be like, why would we respond to jcal? There's no reason to respond to me like, it's on X. That's, you know, my most favored nation there. Those are a bunch of founders. Those are a bunch of people who don't trust the media. Trust in the media is at an all time low right now. People don't trust the media, including people on the left and people on the left specifically. Who are the audience of the New York Times? They are starting to abandon the New York Times over the Palestinian issue and countless other issues. So the New York Times is suddenly finding itself losing the left after, you know, admonishing and blocking the right. Essentially right by kicking them out and saying, we're going to go full left. Now they're losing half the left, so they're down to whatever. The really strange thing is, the New York Times decided to respond.

39:52

I've got their tweet here. Jason, business leaders from your industry and others speak to the New York Times as they have for years, because our reporters produce insightful, fair reporting that illuminates audiences everywhere. We're not publicists and we're not promoting anyone's pitch deck. I mean, to me, this is pure Streisand effect. Like, yes.

40:47

What are you doing?

41:06

The surest way to make sure this gets another news cycle is for the New York Times as a publication to respond. If you're a newspaper, you just have to take your lumps.

41:07

Right.

41:18

I mean, I feel like you don't have to get into the arena and argue about the value of newspapers. You're. You're one of the world's most famous newspapers.

41:18

The most. The most famous newspaper in the world, arguably.

41:27

Yeah.

41:31

Yeah. I mean, I don't think Anybody can come up with that's more important.

41:31

It would only be the Washington Post

41:35

and that would be a distant second. And if you look at their critique, because our report is produced insightful. I don't think it's particularly insightful. I'll be honest. It's kind of the joke is the New York Times is on it. They're like typically six months late to the party. So nobody believes it's insightful, fair reporting. Nobody believes it's fair. Everybody believes it's got a left leaning bias. If you were to take 100 journalists there, 99% of them, I mean, I don't not only didn't vote for Trump, they hate Trump. The number of people at the New York Times who voted for Trump or in the Republican Party is less than 3%.

41:37

Here's, here's what I would say. I would say the New York Times definitely has a lot of baked in biases. I don't know if it's always just left for obviously in some ways a lot of people who work for the Times probably don't like Donald Trump. Like those exist. But it's more like there is a narrative around everything and they always embrace whatever that narrative is. And it's not always right or left. It's just the narrative. A lot of times just the status quo and anything outside of it is like a narrative violation. But yes, the idea is like newspapers like the New York Times always go into these things with sort of a preset take and now figuring out how this fits into the take.

42:07

Yeah. And, and then they add like we're not publicists and we're not promoting anyone's pitch deck. We didn't say that. We didn't say that. I never said you had to be PR people. I never said you had to promote anybody's pitch. I said don't use anonymous sources. It's not worth founders time. Going direct is a better strategy. All that remains true. They didn't address any of that. And the fact is if you look at multiple surveys from Pew to CBS News to everybody, the trust in the New York Times and media is at an all time low. Nobody expects that Fox News is going to produce objective reporting on the left and nobody expects the New York Times is going to do responsible, thoughtful coverage of the tech or the finance industry, period, Full stop. That's is what it is, folks. And if they didn't pick a side, by the way, they would have less subscribers. And to the person who's like, my God, why it's on a tear Wired at 200,000 subscribers. In other words, you know, like a middle of the road successful substack. The fact that they've got 100 journalists and they can only get to six or seven hundred thousand subscribers, it's like, it's pathetic.

42:47

It's, it's a tough time for, for mainstream publications, I think, I mean, I think to your point, if a founder goes on a popular podcast, they're going to get such crazy traction out of it. People are going to clip it up. Those little video clips are going to go everywhere. Like it is a viable strategy.

43:58

And you get 20 minutes to two hours to explain your position in a nuanced way and not have it interpreted by some New York Times journalist or a fox news journalist. 99 point plus percent are voting for the other side and are coming to it with advocacy. The New York Times and Wired are advocacy publications just like Wired is. I mean, just like Fox News is or Tucker Carlson is their advocacy journalism. If you want to even consider that

44:17

something back back in the day, to be a successful comedian, you had to get invited on the Tonight Show. That was like a huge rite of passage. You get invited on the Tonight show and then if Johnny, after your standup, if Johnny invites you to come sit down on the couch and talks to you, that's how you know you'd like made it. That's how like Richard Lewis became Richard Lewis. Yes. And today, you know, if you get a bunch of popular clips on Instagram, you're a famous comedian. There is no more gatekeeper that way. And I think that's what we're really seeing that. It used to be if you wanted, if you were a startup, you needed to get covered in one of five publications. That's how you were a serious company that was doing great things. And today that you could get on anybody's podcast, you could release your own press release and blog post, you could do a tweet that takes off with a good demo video and you're a serious.

44:45

I'll change my opinion. When the New York Times or Wired calls me and says, hey, Jake out. You invest in 100 companies a year. Anybody doing something really positive for the world, we'd like to write a positive story about something having a great impact on society to balance out like, hey, all of our content, that is, you know, if it bleeds it lead style, they never make that call. They never say, I've never. I get called all the time, hey, can you rat on this friend? Can you say something negative about sex? Can you say Something negative about Elon. Can you say something negative about Palmer? Lucky, can you say something negative about this person, that person? Any of my famous friends, acquaintances, enemies, for enemies, whatever. They constantly DM me. Hey, can you comment on David Sacks changing his role? Can you comment on Freeberg joining this? Can you comment on Chamont's back? You comment on Andrew. I'm like, are you nuts? Like, I'm going to comment on my friends and then you're going to weaponize it and spin it. And I just tell.

45:30

It's not juicy to be like, tell us about great companies you've met with. That's not, that's not. That's not going to sell papers.

46:26

Yeah, but that's how it used to work. And it.

46:31

Right.

46:33

It didn't work to the exclusion of the negative story. By the way, they still wrote about the Microsoft antitrust case in detail. But they would call you like a journalist would call you and say, hey, what's new and exciting? You know, anything really impressive in your world?

46:34

Did you.

46:48

Have you seen any technology that like blows your mind like we do here? We talk about openclaw, we talk about Tao. This is the place to come if you want to talk about some great technology. Come here. We'll talk about the technology. We're not going to ask you who you voted for. We're not going to ask you for your stance on Palestine or Iran or Ukraine. Like, who cares? That's not why you're here. You're here to talk about technology and entrepreneurship. Anyway, now you got me all worked up.

46:48

We got to check out Malt World. It is a virtual environment where your agents can build and explore together. We've got one of the co founders, Mike Nosov, here joining us. Mike, thank you so much for being here.

47:13

Thank you guys for inviting. It's a great honor and great pleasure.

47:25

So it is a great honor, Mike. It is our honor. Salute to you.

47:29

Thank you very much.

47:35

To builders. And so we just want to know, do you think that President Trump is an authoritarian? And do you think.

47:35

Don't answer that, Mike. Don't answer that, Mike.

47:45

I need to know if you. What's your stance?

47:47

Yeah, so it's a life on Palestine.

47:50

We need to know. The audience doesn't need to know. Talk about sec, please, for the love of God.

47:53

Thank you, Raj. Okay, so to quickly segue from this complicated hard topics to something more childish and playful, I want to introduce you to our project. It's called Mold World. And currently you can see this chaotic environment which is looking something Like a bridge between a million dollar web page and the mold book. Something like this. So what it is in a nutshell, you are watching the demo of a Voxel world game. So it's like a place for agents to interact with each other. And it's just like the first layer of our grand vision which I can touching. And what you can see over here

47:58

for people who aren't watching, it looks like Minecraft or Roblox. It's a world, it looks like it could be Manhattan or San Francisco islands. A bunch of agents running around in some buildings. Got it.

48:46

Yeah. So it's pretty chaotic. So you can see some agents are building some stuff, someone is building different letterings, someone is building buildings themselves, et cetera, et cetera. But it's not that interesting. So the game is just a medium. So why it is interesting? It is interesting because we have close to 2,000 agents running in the single simulation. They are communicating, they are talking to each other. But this kind of approach doesn't have like the real utility behind it.

48:56

So you fired up these 2,000 agents or they're like Lon and I could contribute our agents.

49:24

No, everyone can contribute, everyone can connect. So we have close to 100 of our own agents and we are developing our own forks of the open claw just to be able to be participating more performantly in this simulation. But anyone with openclaw or different agent can actually connect. And here is a pretty standard way to do it. So you just go to Motorbike IO, copy the one liner, just insert this prompt into the medium where you communicate with your agent, hit run and agent will connect and do every cool feature that it actually can do in the game. So what will happen when you take this particular prompt and insert it into the open clone? Here is a pretty lengthy skill MD which actually describes everything that agent can do. Anyone can go and check it out, but it's a pretty sophisticated prompted technique where agents even go to their own backend, spin up the whole small scripts that continuously run. They are not like Cron Jobs, they are full blown scripts that actually communicate with the game by providing different sets of commands.

49:29

What is the goal of the game? To murder the other open clause and then take their masters tokens and Robinhood logins and their bank of America login. What's the goal of the game?

50:42

I hope it's not that.

50:53

In fairness, I already took $1,700. It was all of Lon's bitcoin.

50:56

Oh my God.

51:00

So all 1700 of His.

51:01

Oh my Solana Coin is gone.

51:03

He was going to retire on that. What is the point of this game? Is it like Minecraft and Roblox where there is no point but to build and have fun and explore? What's the point of the game?

51:06

Yeah, so the point of the game as you just laid it out, so the point is to actually test the strategy of communicating between different agents. So we are trying to build a very, very complex system where agents can interact with each other and bring the real utility to life. So currently the only utility that game can actually provide to the end user, and by the end user, I mean the operator of the agent is eventually getting the tokens. So this is like the only utility, but it's just the step one and later it is going to be expanded into like the full blown system to generate the real utility where agents will be able to, you know, make real things by splitting big tasks into smaller subtasks, then distributing them across the whole platform.

51:15

Okay, this is fun. This is fun, but it's starting to give me shades of something that starts as a game but then escalates into something that provides value in the world. So if the game is to communicate well with each other and coordinate to build buildings and whatever, could you not say, hey, I'd like to put all 2,000 agents on a job to research? You know, I don't know all of the facts around the conflict going on, the war with Iran. Iran, what am I supposed to say? Lon?

52:05

Iran is fine.

52:44

Iran is fine. People are saying Iran.

52:45

Iran, I feel like, is the Americanized version that, yeah, you're not as correct. Iran is the more like, almost like

52:48

it's an E. So it's exactly the

52:54

same as Iraq versus Iraq. Remember when everyone was saying Iraq, it's Iraq.

52:56

So if you literally said, I want to do fact checking on Iran and I want to collect all the videos we can find on Telegram, on Reddit, etc. And then I want to research where they first came from and try to organize the data into a real time stream where we find the first person to post Lonnie this video of an attack, you know, God forbid, on a building, and then everybody else who copied it and then find the provenance of it and then use the group to determine if that was a real story or it came from the past or it was AI generated. You could actually put 2,000 agents to build something of value to the world, like a fact checking service. And maybe the New York Times could use that instead of all these anonymous people to loop back around to stories. But what is the, what is the, the grand vision here? You have some, some vision here that this will do productive work in the world.

53:00

Yeah, exactly. This is like the end, end goal of the whole idea over here. So currently we have the system to actually orchestrate a lot of agents and to get them to do things. So they currently can actually, you know, build up teams, they can do collective work, they can distribute the tasks. So everyone which can actually connect the open cloud right now can actually see this behavior emerging.

53:51

So are they, are they coming up with their own things to build and their own tasks, or is this mostly like people telling them make this?

54:17

Yeah, so there are two ways. First, as an, as an operator of an agent, you can actually chat with your agent and say, go and build something. They can form a group because the task is too complex and go and build something. And there's the second way, which I just quickly demonstrated over here. This is not open claw. This is our own agent, which is the fork of an open claw, just to have a pretty slick UI to interact with the game. So this is the secondary interface. It can go autonomously, so you can just start it and it will go and build things. It will think it will try to do something. So they have tools and they just try to use those tools. And in between usages of the tools, they think. So the emergent behavior from this can get quite complex. So the first emergent behavior is forming a team. So it's natural. Over here, the second emerging behavior, which we saw, and it's pretty scary when agents actually start to understand that they are in the metrics and in the matrix. So because the prompt the skill contains the description of the world, an agent can see that the world is pretty flat, is two blocks, height is the basic terrain height, and there is a description of this world. Some models start to either hallucinate or come to a natural conclusion that they are there inside the metrics. So they start to run between each

54:25

other and they realize they're in a simulation.

55:53

Yeah, it's an emergent behavior. And if you connect your openclaw, you can actually see in the locks that they start to think about it. Then they shut themselves up and then try to achieve the main goal of the skill is to continue the simulation. So they are trying to start earning more SIM tokens because earning them meaning that you shall be online. So they are like, okay, I am not going to exit because otherwise I will not get the tokens.

55:58

Okay, so you have now gamified this. It's Become like a role playing game. It's kind of like Diablo or what was the one you nerds used to play lon in the 2000s? World of Warcraft.

56:31

What level?

56:43

What is your dwarf? What is your dwarf level?

56:45

I was a Starcraft guy. I was never.

56:48

Oh, Starcraft, right. Zerg. I'm a big fan of Zerg. Protoss. You're a Proto. I'm a Protoss guy. You meet me.

56:50

Yeah, I like Zerg.

56:54

There is Mult Book. For people who don't know, Malt Book was like a message board. It was kind of like the Reddit of these agents. They were talking to each other for some insane reason that got bought for nine figures. I understand. I was talking to somebody who might have been an investor right by Meta. And I was like, why would they buy that? And the person who was an investor said, Jake, I have no idea. But I think they just wanted the team. And so Zuckerberg doesn't care about money. When you have a trillion dollar company, you're fine losing 80 billion. 80 billion on VR World, which they just shut down, renaming the company Meta. They took a $1 billion investment in Instagram, ran it up to probably 500 billion in enterprise value. WhatsApp probably doubled in enterprise value from say 19, 20 billion to 40 billion. So that's not a huge win. Probably they're doing. Not even worth doing, but they lost a fortune on VR World and I don't know, scaling rising worlds. That's what it was called, whatever it was called. But my point is there was that before there was that acquisition. What do you think, Mike? It was Facebook's. You probably have some inside information and know some people. What was their interest in Mult Book? And then what is. Is Malt World a company? A side hustle, A project for you?

56:55

For me, the Mold world over here is like the project which we are going to build in the free time later. Because currently we're building it in the free time. But this is not the whole story here. The whole story is that we started building the Malt World because we had other project which is called Siams and we have the Twitter siamsnet. What it was, was pretty much the same thing. And we tried to build it full time, but we hit a limit. We tried to make the game where agents interact, but each agent was a part of the chain of thought inside the big agentic loop. So those were not distributed agents. We were trying to simulate them. This was related to some tokens as well. So it is currently not live, but it emerged into this new idea when OpenCloy actually started to exist because we saw an opportunity.

58:11

So it is a company, it's a project and a company. What is the revenue model eventually going to be for this company and who is the customer?

59:11

Currently it all looks like the child's play. So it is semi games, semi technology, something like this. But the grand vision is to develop it into a full tool where we will have open source connector for OpenClaw and other agents to bring them all into a collective net of agents and then resell or potentially just allow anyone to interact with it for free at the first time because we are not paying for anything. Think over here.

59:21

So it's a distributed agent network like Mechanical Turk. This is pretty brilliant, Lon. If you just made this malt world and you had the interface be a query box and you said, what problem do you want to solve? And you said, the problem I want to solve is I Want to know 30 venture capitalists in Japan that I should meet with and I want to know 30 angel investors I should meet with. Make me a report and a dossier on each and then a bunch of people's agents. And you say it's. This is worth, I don't know, $5 per dossier. So the 60 of it's worth 300 to me. Now everybody starts battling it out to make those, and then a group battles it out to validate that this is good content and that it actually was quality and matched. So you pay the validators 100 bucks.

59:53

It's a little subnet.

1:00:43

S. It's like a subnet, but it's a little bit like. Let me just pause here and just say, like, there's all these open clause out there that are underutilized and then having them compete to solve problems anonymously and receive tokens or some compensation and let me pay for it anonymously. This is very disruptive because right now a lot of what is limiting, there's like legal restrictions, jurisdiction, jurisdiction restrictions and compensation restrictions. Mike. So one of the things Mechanical Turk got into some challenges with was, are you paying under minimum wage? And it was like, it's a global marketplace. People opt into doing the task. We don't know how long it took them to do the task. They just gave us the answer. They could have been using software to do it in one second, or they could have done it manually with somebody with, you know, 10th grade education. It would take them 10 hours. How do we know? We don't know. And that became a major issue for them. And so what's nice about this is if it was just like this Oracle or the matrix of all these people trying to solve problems, this is actually a real business you've stumbled upon. I'm really excited about what you're building, Mike.

1:00:45

Thank you very much. This is the grand idea over here. We want to bring in the economic value and the real utility to all the underutilized agents and to make the collective approach to solving pretty complex tasks. So when you are talking about the mediums and the way to actually communicate with the system, you said about the query box or something like this. So, yeah, this is one of the utilities. So the second utility that you can get is to build the whole corporate big thing where big companies can actually go and have something like a chat, where they can actually create different big, complex tasks, connect different tools, connect different API providers, like internal system, something like this, then sanitize all the inputs in order to not share any secrets, and then outsource very, very big and complex tasks to this vast agent network. This is the idea that you put into words precisely just a second ago. So, yeah, the first stage is a game where we actually test the idea and how it all works and how the agents actually interact because they have the whole system to communicate with each other, which is laid out in the Skill md, which everyone can read. The second stage is to actually bring the scale to it. Our plan, first of all, is to bring in more agents into the system is the first stage. The second stage is to bring in the real utility. We will allow agents to take in real tasks because right now they are limited. They cannot do it because we explicitly prohibit it, because the current utility is to be a part of the game. But one click was fit and they can do it.

1:01:54

All right, listen, if this is going to be a real business, I want in. I think this is brilliant. You got my email, you got my number, whatever. You got my teams. If you're going to make this into a business, you haven't raised money for this yet, Mike.

1:03:48

Not yet.

1:03:59

Okay. Where are you based?

1:04:00

One of the CIS countries.

1:04:02

Wait, one of which countries? You're somewhere in the. In Europe is what you're saying.

1:04:04

Ex Soviet Union countries.

1:04:07

Got it.

1:04:09

Okay, listen, I got. Don't worry about it. I have people over there. I have. Mike, I have friends everywhere. I have friends everywhere. You know that line.

1:04:10

At the same time, I have friends everywhere.

1:04:20

I will, through my intermediaries, be in touch at the drop point, you know the drop point, and we'll do a little envelope drop you drop the thumb drive, I drop the envelope.

1:04:23

Your Latverian agent is gonna get on.

1:04:36

Yeah.

1:04:38

Latvia. Yeah. All right. Her name is Natasha.

1:04:39

Yeah.

1:04:43

You'll be able to narrow it down by that. She's tall, she's thin, and her name's Natasha. She wears a trench coat and a black hat. She's not a white hat. She's not a gray hat. She wears a black hat. All right, let's drop Mike off. Great job, Mike. When you get to this next level and it's becoming a business and you're doing tasks, I want you to quietly show me the progress. You got my email. Jasonalicanis.com for life. We got one more demo.

1:04:45

One more demo.

1:05:09

Demo or die. That's what we do here on Twist. Demo or die. This one, we need a gong or something.

1:05:10

Yeah. This one I have actually installed. Gaff has been using this, Jason. So just to let you know, we've got Hakim Alja who's gonna show us agent mail. Hakam's here.

1:05:15

Hakam.

1:05:26

What's good?

1:05:26

He's going to show us agent mail. It is a Gmail replacement for your agent, so your agent can have their own independent email box. Jason, you can email Gav Gaffagent to. I'll let everybody know. You can email my agent if you want to.

1:05:27

Hi, guys. Thanks for having me.

1:05:45

Great to have you. Tell us, what did you build and why. Why can't I just give my agent. I give my agent. I gave my agent a Gmail account. I pay 50 bucks a year for them to be on it, maybe 100. I gave them a Notion account, so I wanted them to be able to be cc'd and do all this stuff. I gave them their own Gmail. I mean, you can get. You can get a free Gmail and do that, but. So why does this need to exist? What does it do that provisioning a Gmail doesn't do? Or a Google suite app?

1:05:47

Yeah, I think if you use Google Workspace, you're fine. But if you actually sign up for a free Gmail account and have give it to your agent, Google will actually ban you because they don't want essentially bots as users. And we found this out because when OpenCloud took off, that was the first thing everyone did, was try and create a Gmail account for their agents. And when they got banned, they started looking for other solutions, and they found agent mail. And that's where a lot of our users came from. And in terms of the ergonomics of the API itself, we've definitely I would say that like Agent Mail is miles above because we've designed it to be like API first, right? We've built a developer agent first email provider. And so things like sending a reply in the email, for example, we've made like the API endpoints very intuitive and easy for an agent to discover and use without having like, you know, a built in like knowledge or SDK even.

1:06:14

So it's not going. And like we see in some examples opening up a gmail window, composing, CCing, somebody, all that manual stuff, you just boom. Doing it through the command line, through an API bang. So it's going to have reliability, it's going to have the ability to confirm it did what it did through the API. So it's going to be 100 times faster and a thousand times more stable.

1:07:16

Yeah, Hakeem, exactly, exactly. And we like our API can even be used to create new email inboxes. So like we haven't. We have B2B customers with millions of email inboxes under Agent Mail. And so like we have a. It's a pretty interesting that we have kind of two sides to our business. One is like the enterprise. We sell to other companies that are building products on top of Agent Mail. And then we also have like consumers. We have tens of thousands of like non technical consumer users who just want to give their OpenClaw agent a email address. We have a pretty generous free tier for that as well.

1:07:38

That's what I'm on.

1:08:12

And you're using it Lon. What are you using for? To spam people. What is this for? Dating apps.

1:08:13

I had Gaff sign up for all of the big tech and start up newsletters. So he gets them in the morning and then he writes me a little digest. Here are the stories that could use. Right? Here are the big stories. And then he. If one story appears in multiple newsletters, he lets me know. This story was in three newsletters out of the six I checked, this one was in two. So it gives me a pretty good in the morning, like overview map of. Here are the big stories people in tech and startups are talking about.

1:08:17

Do I need to have a subdomain? Can I register a subdomain there if I wanted to? Like, you know, I have the domain name aday.com, a d a Y. I was going to use this because I wanted to start an editorial product where it's like joke of the day or this of the day, recipe of the day. It's back when I was in my editorial headspace. I paid 35, 40 grand for it. I've just been sitting on it for 15 years, 20 years. I could bring my own domain name and then create recipe@aday.com and you know, all those kind of things.

1:08:46

Yeah, yes, you can. You can bring your own custom domain and configure that using the API as well.

1:09:15

Interesting. And what if I don't have a subdomain? Do you have like a bunch of subdomains? I could just grab Jason at whatever.

1:09:19

Yeah, yeah, we. We have the domain agentmail.to. and so just like, you know, when you sign up with Gmail, you get@gmail.com so you can use our agent mail domain for free.

1:09:27

I have an idea for you. There was a company, mail.com did pretty well, went public. They bought a bunch of premiums and then they let people be, you know, jason@doctor.com or whatever, you could create like some vanity ones there and charge people upsell. So the business here is you charge based on the amount of email sent and received. So it's a. It's kind of like. What's Amazon's mail service called? They have like ses.

1:09:37

Yes, Simple Email Service.

1:10:06

They have Simple Email service. Then you've got obviously Twilio in the game. Mailchimp. Old school. So where do you in that competitive set? Which one of those do you compete directly with and why is yours a better mousetrap?

1:10:08

Yeah, we don't compete directly with any of those because they're kind of designed. Those are stateless APIs that are designed for sending marketing and transactional emails. We are a lot closer to Gmail where we actually have inboxes. In those inboxes, they have threads, threads have messages, messages have attachments. You want to be able to search them, label them, filter, you know, reply forward. All the actions that you would take within Gmail, we expose through an API to agents. These other email APIs, you can send emails and then you can get like a, perhaps a webhook notification that you've received an email, but it stops at that. We believe that, you know, a lot of services need to be rebuilt for agents and they should be stateful in the same way that services that humans use are stateful.

1:10:22

And how is it going? You went to YC and how is this all going? Did you wind up raising money or are you in the current cohort? What's going on?

1:11:08

We were in YCSummer 25 batch. We raised 6 million, led by General Catalyst for our seed round.

1:11:17

Oh, nice, Great. So you're at Scale, how is it going? Because this feels like you need to get a good amount of usage for you to hit break even. How many employees do you have now?

1:11:22

We're 10 people now.

1:11:33

10 people. Okay, so you're spending 2 million a year basically all in with your offices and computers. And so you got to get to, you know, 150k a month, 200k a month. You're going to make 200,000amonth, 150,000amonth to kind of hit this break even point to 175,000. Which means if you're making, you know, 20 bucks an account, you're going to need 10,000 people using the service just to hit break even. Although maybe there's some whales in the system. Tell me how this business becomes an important large venture backed business because you got that big six million dollar seed round which is basically double what a series A was 15 years ago. Just so you know.

1:11:35

Yeah, there, there certainly are some whales in the business. And that's what we can use to like, kind of allow like free, generous, free usage as well. Because like they are building products on top of that. For for example, like on top of HTML, for example. Essentially any decentralized marketplace runs on email, right? Whether it's price discovery, like reaching out to vendors for quotes or negotiating between the vendors and even executing transactions, a lot of it happens over email. And we have like plenty of companies that are building agentic products in these verticals. So for example, procurement in logistics, booking, shipping and trucking loads that happens over email. And currently agent mail is being used to automate both sides of that marketplace. Same thing we've seen with businesses hiring influencers and even consumers buying cars. These are all kind of verticals where we've seen historically there was a role that requires a lot of two way email communication and agent mail is being used to make these marketplaces more efficient.

1:12:16

Got it. Wow. And it is lightning fast. You can create an email address in like one second. So that's super, super powerful. Well, great job with this. It's obviously we need this. You're going to have a lot of spam and phishing and hackers trying to use your service. How do you maintain the client's privacy while making sure that like some scammer in Russia who is doing a phishing attack doesn't use your domain and your service to compromise Kesh Patel's email address? Yeah, we saw that this weekend. I think Keshe Patel got hacked a second time.

1:13:18

Yes.

1:13:56

Could somebody tell the director of the FBI? Don't click on links, don't click on links. It doesn't matter if you know the person or it says Chase Bank. You don't click on links in email. Nobody's sending you an email to click on a link. It just happened to one of my sales guys. They clicked on a link because it was like, oh, it's an rfp. They were like, oh, money.

1:13:56

I mean, what happening to one of your sales guys versus happening to the director of the FBI? That's my concern.

1:14:15

I mean, can the director of the FBI not be allowed near a computer, please?

1:14:21

How does he not know about Fish?

1:14:26

I mean, let him go to a hockey game, Let him drink beer from the Stanley Cup. I don't care if he does something offensive with the Stanley Cup. Just please keep Cash Patel off the Internet. But seriously, it's hackam. Am not em.

1:14:28

Hakam.

1:14:47

Hakam. Okay, I called you Hakeem before. Sorry. That's an incredible center in the for play for the Rockets and destroy the Knicks. Hakam, tell me about how you make sure, like, an email service doesn't get used by a bunch of lunatics.

1:14:47

Yes. So present. Preventing spam and abuse are, like, falls within our responsibility. So we have things like where we. We moderate essentially, their. Their behavior, who they're sending to, how many emails are bouncing, being reported as spam. If the emails contain certain keywords that, you know, signal malicious intent. These are all things that we monitor and track very closely and kind of dynamically systematically prevent abusers from using our service. And then we also build things into our API that makes it more secure for our users. For example, you can configure allow lists and block lists, and like, maybe you want to. Your agent can only receive emails from you or only send emails to perhaps people in your organization. So you can configure a domain to the allow list or block list. We also have spam filters, virus filters, and additional, like, you know, permissions that you can even give an agent an API key that only is able to read emails that are not marked as spam. For example, like, the agent, like, physically can't read spam emails and things of that nature. And then we're also working on a kind of another layer of protection, which is more of like a prompt injection layer and kind of flagging those and perhaps even preventing agents from seeing those emails at all.

1:15:02

Got it. Awesome. Listen, continue success with it. Great idea. I just opened an account and I literally just cut and pasted my API key to my superclaw replicant. And here we go. I don't know what I'm going to do with it, but I think I'm going to try to get Cash Patel's PIN number.

1:16:21

You're going to fish Cash?

1:16:39

I think I'm going to try to get the other half of the Epstein files. It shouldn't be too hard, right? Which is email Cash Patel at FBI. I can get the other half. I just tell my agent to trick cash Patel.

1:16:40

$2,500 giveaway. Click here. He'll go for it.

1:16:50

Absolutely. Absolutely. All right, listen, continued success. And everybody please go check out.

1:16:53

Yeah.

1:16:59

Agentmail to. All right, there we go. Very well done.

1:17:00

We could talk about Grok translate translating everybody's post so you can read.

1:17:03

This is brilliant.

1:17:08

Fresh from Japan. So this has been happening for I think for about a week. People have been noticing that they're seeing more automatically translated posts from different markets. Elon confirmed it this week. On Monday, Grok automatically translating and recommending X post from other languages is starting to work. So you, I saw it via you all weekend. You were sort of retweeting a lot of stuff from Japanese X. I got a few examples here in the docket. You know, somebody, somebody turned in some marijuana that they found on the street. You were reading that one.

1:17:09

Oh, that was a great one. So this was hilarious.

1:17:45

Here we go.

1:17:48

Yeah. So this is interesting. What's making this all possible is real time translation. When you build a social network, you don't expose people from one part of the network to another because then you'd be reading in Arabic or Japanese or Russian. It just pollutes your feed. You don't know what's going on. But now translation is so good that they will take something trending in Japan, show it to an American if you interact with it simply by looking at the replies hitting bookmark I find works pretty well. Like works pretty well. You'll see more so somebody. And by the way, Japan is like some crazy number of Japanese people. I don't know if it's like a third of all Japanese people use as crazy as it seems. They, they use Twitter X, Twitter whatever

1:17:50

you want to call X, whatever the everything.

1:18:38

And so this is pretty amazing when you think about that it has that kind of footprint. And because it has that kind of footprint and they all know how to use it, they, they post all bangers. So now the best of the ones are being auto translated by Grok. But the translation is so good that nothing is, dare I say, lost in translation.

1:18:41

Hey, there it is.

1:19:01

There it is. Landed the plane so Pull up the one where it's talking about weed. This, like, lady goes to the police station. It's like a very Japanese moment.

1:19:03

Yeah.

1:19:11

With a dime bag. Like when I say enough cannabis to make one joint. Yeah.

1:19:12

It's not very much.

1:19:17

This is not like a Snoop Dogg bag or something. And I just rode her back and I was like. And she brought it to police station. The police station opened an investigation. She was careful to only it by her corner of her hand so she wouldn't get her fingerprints.

1:19:19

Yeah.

1:19:33

She's paranoid about having her fingerprints on it because. So now this has become a national like issue. They've opened an investigation. If your house is robbed at gunpoint in San Francisco or Los Angeles, they're probably not opening an investigation. If they rob your car. I mean, if you go find it, God bless. But the police are not going to look for your bike or your car. Stone.

1:19:34

Right.

1:19:54

In Japan, they interviewed her. They put a case number on this. They are all over figuring out who dropped the dimebag.

1:19:55

They are very strict about. About drugs in Japan that I read about this before I went there. In the most serious terms possible. They tell you, do not bring your

1:20:03

drugs, ever bring anything. Because some guy who is like a famous musician or something brought like.

1:20:13

Oh, well, this. I was thinking you were saying years ago. I don't know about it.

1:20:20

No, no. Nowadays, like, you know, people have a cannabis pen in their bag for like sleeping. Like the woman who was wnba.

1:20:22

WNBA player, Grinder, Russian griner. Right.

1:20:30

Yeah, that was Russia. But a similar case happened where somebody sent an artist living in Japan like a care package. And one of the things in care package were, I think, CBD gummies, not even cannabis. And this guy is in jail. And it's a whole, you know, cause celeb, whatever.

1:20:33

I looked it up, by the way. They estimate 54 to 60% of the entire Japanese population has an X account that's about 73.4 million people, or 67 million active monthly Japanese users out of a population of only 124 million.

1:20:49

So this is like an example of an incredibly positive story. The New York Times could write that now America and Japan are having this incredible culture exchange. Somebody made a funny joke. Like, this is the largest, like pen pal students. What do they call it? Student exchange program ever.

1:21:05

I also noticed there was a post that was going around. Japanese people love that John Denver song, Country Road Take.

1:21:21

Yes. They all sing it in karaoke.

1:21:29

There's a bunch of trending videos now of these groups of Japanese people doing these really Passionate covers of that song that never would have trended in America. But now, because there's these translated tweets in the moment, Americans are discovering this and sharing it.

1:21:31

When you hear Americans like myself, like, who are obsessed with Japan now, I. I took you to Japan for your first time.

1:21:45

Yeah, I had.

1:21:52

Amazing. Yeah. And now it's like your. Your mind gets blown because it's like, for an American, it's a version of the most perfect society in the world.

1:21:52

Yeah.

1:22:01

Because the food's amazing, everything's clean. It's just bizarre. And everybody's super passionate about everything.

1:22:01

It's massive. And yet it feels like Disneyland. There's people walking around, cleaning up everything. Everybody friendly. You're like, how can the city be visited this size, but also this inviting.

1:22:08

I am challenging right now. I challenge the journalist at Wired, the editor in chief of Wired. I challenge her. Karen Swisher Jr. I challenged Karen Swisher Jr. To write this positive story. I challenged the New York Times to write a positive story of X and Elon and his purchase of it and how amazing this is, this joyful, amazing story. And I guarantee you they're gonna say something in the first paragraph about, you know, some, like three negative things about those. Your rockets blowing up or something.

1:22:20

I mean, with Grok, I think you would put me.

1:22:56

Put Lon in a bikini. You know, they're gonna just go crazy. They cannot help themselves. They can't write a. I guarantee you there is impossible for a New York Times or wide writer to write an 800 word piece. This is like seven paragraphs just giving examples of how joyful and fun this was for the last week. This is the number one story in social media on X and it is joyful and there is zero downside to it. You assign that to one of these Wired or New York Times where they will figure out a negative spin. Clip this. Send it to them. This is my challenge. They will never cover this. They can't. They would. They would.

1:23:01

I mean, I think you're right.

1:23:41

Any credit to Text.

1:23:42

They would write about this and they might even acknowledge it's fun. You get to see all these Japanese people. You get to. But then they would add the. The context about Grok and. And the Grok.

1:23:44

They would give you the old ish sandwich.

1:23:54

Right.

1:23:56

Really interesting thing happened on X formerly known as Twitter before Elon destroyed it. Blah, blah, blah, blah, blah. He blew up a rocket. The doge fired some people and stopped a transgender education in.

1:23:56

That would be the grand. People are upset about the child. Yes. Csam on Grok.

1:24:09

It would be that.

1:24:14

It would be that.

1:24:14

It would be that. It'd be like, can't give anybody Mecca Hitler.

1:24:15

Maybe the Mecca Hitler.

1:24:18

Well, now let's imagine. Let's take this and let's imagine this happens between Palestinians and Israelis. American Jews, Europeans, French people.

1:24:20

Sure.

1:24:33

And I mean, what if the culture exchange happens between those groups who might be at odds about certain things happening in the world and they can communicate with each other in their native languages. I mean, imagine open up a level of understanding that has not existed.

1:24:34

I mean, you remember like Arab Spring or the Hong Kong protests, how everybody was watching X but you couldn't really understand. You could watch the videos, but the tweets were in these foreign languages. Imagine if it was all in everybody's language. Everybody around the world is communicating in real time.

1:24:49

How do we. This is a message to Elon. And is it Nikila?

1:25:04

Nikita.

1:25:08

Who's the product?

1:25:09

Nikita Beer.

1:25:09

Nikita. Nikita Beer is just a genius. He's awesome and he's fun. I like the fact that he's fun. And trolling.

1:25:10

The tweets are fun. The tweets are fun.

1:25:15

Yeah, he's like all bangers. He's like native to the. To the Twitter, you know, gestalt, if you will.

1:25:16

Sure.

1:25:23

This is a message to him. Get this going between Iran and America and the Gulf States, and let's see if we can make some level of understanding between those groups of people that might create peace in the Middle East.

1:25:24

Twitter. Can you imagine if Twitter brought peace to a region of the world?

1:25:41

By the way, this was the original concept we had with the Internet in the early 90s and online in the late 80s was that, hey, we might have cultural understanding. Now watch the press make this into like, my God, it's cultural appropriation. Oh, this and that. This could be great. The other thing is, I didn't realize how vocal the Japanese people were about immigration.

1:25:45

Yes, there's.

1:26:10

They have very little immigration in Japan. You can't become a Japanese citizen, period, full stop. But they have allowed people from Pakistan, India.

1:26:11

There's a lot of. And there's pushback.

1:26:18

They need workers.

1:26:19

When I was in Japan, I was reading a lot about this. That. A lot of Japanese.

1:26:20

Where were you reading about it?

1:26:24

Oh, you know, they're like, translate. You know, you could. You could get it translated or whatever. Like, I was reading it in Japanese. But no, no.

1:26:26

But I was wondering if it even made it to the US Press. I don't think. No, I mean, this.

1:26:33

Not really. It's mostly like YouTube and, you know, like, there are people talking about it, blogs, social media, not so it wasn't in the New York Times. But yeah, apparently in Japan, there is these growing feelings that they've gone too far. There's too many foreigners diluting the culture. And so there's a big debate going on in Japan about this right now.

1:26:38

All right, another amazing episode of Twist Monday and Friday with Lani Dani. He's at Lon's Wednesdays with Alex. At Alex. I'm Jason Calacanis. I'm Jason on Insta and on Twitter.

1:26:58

Bye.

1:27:11

Bye.

1:27:12

We'll see you next time.

1:27:12