Live AI Agents Building Session | enso x NFX

本站添加: 2026-05-30

[00:00:03]All right, I think we can kick it off.

[00:00:05]Um Keith, how are you doing?

[00:00:08]>> Yeah, I'm doing great. Um well, thank you for doing this.

[00:00:11]>> Absolutely.

[00:00:12]>> Um I mean, just for um just for a little bit of um context, uh I mean, we I don't work super closely with Mickey, but you work closely with Giggy and but Giggy is um always singing your praises about how you're embracing AI agents and >> [clears throat] >> you know, I guess one of the kind of the biggest common question we get from portfolio founders is um you know, how are other companies using AI internally kind of from a practical perspective?

[00:00:49]And then also benchmarking themselves to say like, "Look, we've kind of like doubled um you know, output or kind of shrunk these times by you know, like two to five X in the last six months."

[00:01:04]You know, and and sometimes people are like saying it's delighted and other people say, "No, I should be like it should be much more than this." And so, you know, that the idea for this was really just to like um Mickey to kind of share what's um he's been doing. Um and then also just to kind of open it up to kind of share best practices so that um this community can share many of the great things. And it's and as we all know, what is progress today is very different from three months ago and very different from six months ago.

[00:01:37]And so, like this is going to be a very sort of dynamic topic, but it's is probably execution here is probably the kind of single most important startup execution um uh skill that founders have to master right now.

[00:01:56]So, With that, I'll I'll pass it over to Mickey.

[00:02:00]>> Appreciate it, Pete. And you know, I run a small community in Israel of different founders and CEOs and builders.

[00:02:09]And I talk about AI. It's like a WhatsApp group. And and I'm always amazed how I have almost 1,000 people in the group. They get 30 messages every day from me about AI updates and nobody leaves.

[00:02:22]I mean, some of them give up, but but the overall the the the community grows. And and it's very very exciting to see how excited are people about this space and how rapidly it's all changing. So, I think to your point, yeah, it's not every 3 months. It's every 3 3 hours. That's how it feels right now.

[00:02:41]>> Yeah.

[00:02:42]>> But you know, Pete, before I'm kicking it off, I am curious, you know, you're a a GP at at NFX. You've been, you know, building companies. You've been investing in companies.

[00:02:52]You've been doing so so many things.

[00:02:55]Um How do you look at this entire space now and how do you think the ultimate entrepreneur should be utilizing all of those tools?

[00:03:09]>> Uh I guess that Um how do we think about this space?

[00:03:15]Um you know, I'd say the world is quite schizophrenic right now, right? It's like people are kind of unsure what is going to kind of happen in the world in the next couple of years.

[00:03:27]Um And I and I I guess the you know, a lot of you know, conversation with founders, how do I insulate myself against AGI or protect myself against AGI?

[00:03:39]And I think probably the only kind of the only inoculation is to be um highly proficient um with AI tools and moving incredibly quickly.

[00:03:54]Um incredibly quickly. Um uh and so I I and it's you know I I don't think there's any I wouldn't sort of pattern recognize that 19-year-olds are better at this than 39-year-olds.

[00:04:09]Um but um uh I think it's mostly about just high levels of proficiency and high levels of curiosity and high levels of speed. And I think that is really what kind of as we think about as what we're trying to solve for and try to identify.

[00:04:27]>> That's great. I appreciate it. I'll dive into what it actually means behind the scenes to actually build those stuff. Um and then uh people of course if you have any questions uh along the way I'm I'm happy to uh >> Great.

[00:04:42]>> who to address. So, I'm removing you from the spotlight and I'm taking it all for myself. Uh but just to share my screen.

[00:04:50]Um okay. So, hello everyone. Good to see you. We have all types of different um um founders here all the way from the Silicon Valley to Israel. So, this is very very exciting. We do have uh New York as well. So, hello New Yorkers and hello Tel Aviv and hello San Francisco.

[00:05:08]I will kick it off with my you know common session that's called Pimp My Agent. Uh if you were raised or at least cool in the '90s, um then you probably watched the Pimp My Ride uh show. So, that's basically was the reference. Um and I think the most important part of the session is less about the short deck that I'm going to share with you. It's an open session where you can ask questions and we can do stuff live. Okay? So, I am spending um about 19 hours a day doing this. So, for me this is very very close, but the reason I'm doing those type of sessions is because I really want to make sure that I share all those advanced tools and and opportunities and mainly way of thinking with as many people as possible.

[00:05:53]Um, I will introduce myself just for whoever doesn't know who I am. I'm the founder and CEO of And So. And So is an agentic GTM company. We basically help um, startups and and small enterprises uh, use AI agents to find growth hacking opportunities and get customers in very very very very fun ways. I'll present it in a in a second. But, you know, uh, feel free to to reach out if you have any questions around the AI agents. We probably tried all of it.

[00:06:21]So, we'll start from what an AI agent is and there's always a lot of misconception around this term because it's a new term, right? It's not It's not something that we used to um, uh, you know, use um, in our any any of our work. Um, and it basically was created, by the way, in Oxford um, about 40 years ago. It was a new type of automation. I actually met in person the person who invented this term. But, agent by the end of the day or AI agent as we all know it is just software, okay? So, it's just like SaaS or anything that we used to know um, but it just has a new uh, new capabilities and new characteristics. And the main thing is that SaaS was designed for people to use it and this is the first ever software that uses itself to complete tasks.

[00:07:10]Okay? So, that's basically the main difference. So, whenever you're thinking of AI agents, it's It's really important to just remember that there's nothing very special about agents that you don't know about, but there's a just a different concept when you're thinking about it. And the reason this is important is because when you're designing something like this, you can't think about it in the in the same way that you uh, would develop uh, software for people because the you know, machine that going to use it is AI. So, we're going to touch this in a second.

[00:07:40]Um, so the way it works um is that you don't really have to change the way you work certainly, but you do have to change the way you think about um AI. We do have some people here who are not muted. Please make sure that you're muted because we can hear you. Um, so and the first thing I always say is that you have to change your mindset. The reason it's important to change your mindset is because it's extremely hard to pull off all of your knowledge from the previous uh cycle of, you know, tech and just try to kind of build it on top of AI. It's just a different beast. So, when you're thinking about it, you can't just take your patterns and say, "Hey, I'm just building another UI, but now I'm using AI." It's just a completely different type of software. And the reason it's the it's a different type of software is because as I mentioned, we used to create software for people, now we create software for software. That's different. We used to create software for people, so UI, you know, call to actions, design, user experience. All of this is irrelevant when it comes to um AI agents because AI agents basically are using tools, so we'll touch this in a second, to try to achieve a specific task. Now, for example, um monday.com is a great example. It used to be the way for people to manage their work.

[00:09:00]It's still a way for people to manage their work, but in a world where AI is doing things, the question is why do you need a software that manages your work if you have software that does your work? And that's basically the entire pivot of the uh um SaaS companies on the stock market and and why uh stock prices or, you know, company uh uh companies are losing so much of their valuation.

[00:09:24]So, the way we were we used to think about software, so we had two ways to go, right? There's one that is build and and one that is buy. We used to only buy software because it was impossible to build. Okay, so that's something that AI changed rapidly because coding is very much accessible to everyone. It doesn't mean, by the way, everyone are engineers now, but it does mean that everyone can build anyone can build whatever they they think about using, you know, vibe coding tools such as base 44, lovable, whatever. Um so, we're much more inclining now just as companies or as founders to building stuff internally as opposed to buying. But, it's still happening. Okay, so we have a company growing extremely fast because we're selling um off-the-shelf agents um to companies. So, we're still selling software, but we're talking today about building software internally and not uh buying.

[00:10:18]And I want to try to simplify the entire AI agents um uh um landscape for you and to try to kind of under explain what it means in in in what when you're looking at this category. There are two types of AI agents. That's it. You won't really find in 99% of the cases any more types of AI agents actually integrated into production.

[00:10:45]One is called an agentic workflow, which is basically a scripted software just like you would build with, you know, code 5 years ago before AI, but it uses AI to um make certain decisions. So, for example, and we'll touch this in a second, um in on the left side, we will give the agentic workflow the exact prompt or the exact flow of the steps we want it to to do. So, that's a completely different approach from the right side, which will which will touch in a second, where the AI is actually defining the steps. These are the two type of agents we have in the world today. Nobody has ever been able to convince me otherwise. In 99% of the cases, that's what you're going to see in production. The way you make the decision whether you're choosing the right side or the left side, whether you choose the steps or the AI is make is choosing the steps is basically by understanding how convinced you are about the success criterias of the software you're building. So for example, we're now building um a use case in here that scans my email every morning, remove emails that include promotion ad and whatever. So we we want to make we want it to be spam free. And then we ask it not to to delete any emails that are including ansaldo.bot which is our URL. So in this case, I defined what are the steps to be to clean up my my email from uh specifically spam. That can be scripted into 1 2 3 4 5 steps that are using AI. So we're going to touch in a second how we're using AI, but AI is basically just doing the actual work based on my script, based on based on my flow. On the right side, we're using an autonomous agent. And an autonomous agent, it makes the decisions by itself.

[00:12:37]So instead of telling it how to do stuff, we just name the goal. And these are the two type of agents you usually going to use because when you don't know how to reach a specific goal, AI can help you think about those ways and iterate fast to try to get to a to a point where it knows what works best.

[00:12:57]And this is incredible for example, if we're building a software that is trying to do let's say we're doing SDR work. We're trying to do cold outreach.

[00:13:08]And let's say we don't know what is the copy that converts. On the right side, we will tell the AI, you go crazy, come up with ideas and let me know what actually works. On the left side, we'll actually tell the agent, listen, here's what's working for me. I just want to automate the process. So these are the two agents you usually going to see um um in production. There are three layers in agents. And this is extremely important.

[00:13:35]We do have more layers such as, you know, GPUs and and stuff like this, but we're not going to touch this because this is the very much core infrastructure. But we do have three things that we have control over.

[00:13:47]Number one is the LLMs. And we know very modern LLMs or very kind of, you know, famous LLMs that everybody keep talking about such as Claude or OpenAI or Gemini or DeepMind from time to time.

[00:14:02]But there are hundreds of different LLMs. And I'll share a link later on to where to find them that are specialized in specific things or more flexible or less flexible. And the hard part about building an agent is actually finding what's the right LLM to solve the specific use case you're thinking about.

[00:14:20]The second thing is platforms. And platforms are basically where are you going to build your agent? There are, I think, thousands to this at this point of platforms where you can build agents.

[00:14:32]Some of them are more codish. So, that basically means that if you know how to write code and you have experience with engineering, it will give you more flexibility to work the way you want.

[00:14:43]And the other ones are less technical and then you can use use it to prompt those agents. All of them by the end of the day are writing code. So, as I mentioned, AI agents are just software and those platforms know how to how to actually integrate those or to write this code to actually make this agents run.

[00:15:03]The most important part is actually tools. And tools are super super cool because they're basically the way to think about the AI agents as a layer of a new type of internet. And tools are basically what helps AI agents to perform whatever actions they want. So, think about the '90s. You had in in the first browser, and you had the first email, and you had the first wallet, and you had the first news site, and now we're basically rebuilding those tools for AI agents to to use because they have to be, you know, as I mentioned, designed for machines and not for people. So, no UI, headless, and and agents can use it.

[00:15:46]The LLMs, as I mentioned, I'll share I'll share a link later on to open router. You can write them down. That that basically a good way to just discover new LLMs. There are thousands of those. We're exposed to the ones that have most of the money and distribution power, but there are thousands of those, and I really really am encouraging you to guys try it out because there are few cool ones, and you know, Kimmy, for example, is incredible. Deep Seek is pretty cool, and there are few others.

[00:16:13]Uh platforms, this is very important because platforms are basically giving you the ability to just decide what works for you, okay? So, lovable will be great, or any of those live coding tools, Replit, Base, whatever.

[00:16:28]They'll be great because they will give you the ability to write something with UI, and then it just becomes like a SaaS, but it uses, you know, AI to do stuff.

[00:16:38]There are more advanced platforms that are using code such as LangChain or LangSmith in this case on the right side. Those are more for engineers.

[00:16:46]There are more easy ones that I don't have here such as any tool, and there's the the, you know, kind of a drag-and-drop Zapier experience. And there are platforms like Claudy, so claudy.ai. I think I'll I'll share the I'll share all the links later on that can host, you know, autonomous agents and stuff like this. I usually prefer myself to use the most non- flexible, I would I would say, or non-technical platforms just because it just gets extremely hard to to manage and figure out how to use them. So, these are the parts.

[00:17:19]Tools, as I mentioned, are the most important uh part in the ecosystem because your agent is pretty much dumb, or it uses the LLM or whatever it has.

[00:17:28]Um but if you want your agent, for example, to have an inbox, so you can send it an email, you have to have a an infrastructure for this. So, AgentMail provides exactly that. And I'll share my screen for a second. If we go to AgentMail um.tool, this is literally a platform that only creates email inboxes for AI agents. And if you want your AI agent to be able to communicate with you or your customers or whatever over email, you need AgentMail.

[00:17:58]And uh another example here will be uh Browser use. So, Browser use is an incredible company. Uh I think they're based in SF, by the way. Um they give you the ability to just open a browser.

[00:18:11]So, if I am an AI agent and I really want to travel in the web uh through um a virtual browser, this company gives me the the ability to do that. And there are thousands and thousands and thousands and thousands of those um different tools that are basically making your agents more capable of doing stuff.

[00:18:31]Now, I have a great um partnership with NFX on a small microfund that invests specifically in those types of companies, companies that are uh building um agentic tools. And if you go to autonomous.capital.ai and you go to tools, you basically can find all the tools available to AI agents pretty much online. So, this will give you um all the possible abilities to pretty much, you know, equip your AI agent with literally any possible um opportunity. The way to think about it, it's a little bit like skills on on Claude, but it's a little different because it's code that that it's code that is pretty much integrated into um into uh your AI agent.

[00:19:16]So, this is basically the way to think about AI agents. That's it. I wouldn't go into There are many, many different types of AI agents that are experimental, such as agents that are combining more power and less But, whatever. In 99% of the cases, what you want to achieve will be able You will be able to achieve with this type of infrastructure. So, that's what you're thinking about when you're designing your AI agent. Um you should fall in love with one of those platforms and just to iterate on on them. I assume most of you tried, so I assume some of you have already um the opportunity to do so. And then, once you have that, you do have to think and to discover the internet a little bit about with what types of tools are available for your AI agent to achieve the type of things that we care about. So, today we're going to build together um a um an AI agent. I think we'll go with something basic, um unless you want to share an idea, by the way. So, I I can come up with something on the fly.

[00:20:24]So, I'm stopping the sharing the screen for a second, and if you have If someone If someone wants to come up on stage and just ask for an idea that you guys have in mind, I'm happy to share how to build it just on the fly.

[00:20:38]So, think about a problem that you're trying to solve or um a process that you would like to automate.

[00:20:49]>> I I can go. Like, in terms of Thank you, Mickey. In terms of uh you know, I'm thinking about a podcast, and and one of the thing that I have, you know, one agent that is doing the the research and doing kind of like a pre-production. I think that the the thing would be like a post-production. So, getting the you know, the the recording, the transcript, um cutting it to, you know, different according to to to my uh about me and and my tone of voice, cutting the different things, connected with other tools uh that can actually create the videos uh for Instagram, for LinkedIn, and so on.

[00:21:33]Um edit that. Yeah, that's that's one one idea.

[00:21:37]>> It is It is a great use case. I will uh unfortunately have to pass specifically on this because we don't have enough time to do that, but I will iterate on your idea if you don't mind.

[00:21:47]>> Yeah, please.

[00:21:48]>> I'm happy to share I'm happy to share, by the way, how to build stuff stuff like that, but it I it will take a little more than 30 minutes. Um but I'll I'll take another use case. I mean, I have a podcast as well um and um and feel free to listen. It's a with Forbes um as partners. And uh uh one of the things that is extremely hard is to get people on board, right? Because there are always people that you really want to get on board, and then you have to look for those people, find their information, um ping them, um make sure that they have a Calendly or something where they can uh book all of those.

[00:22:22]There's a lot of ping-pong there. I think this will be just easier to build uh when in our time frame. Um Yuval, if you let me uh go for it, I'll we'll do it right now. Perfect. So, this is what we're going to build, okay? And the way I'm going to think about it, I'll basically share my chain of thought, so you have um full accessibility. So, just like you would build any solution on planet Earth, um I would first define the problem. So, the pro- the problem in this case specifically is that um and I'll, you know, kind of go for my problem. I'll say, um it's hard to research who might be relevant for my podcast based on whatever the type of audience I already have plus uh based on um you know, people who are actively interviewing for a podcast because some of them are not. So, I think the problem in this case will be specifically how to find people in the AI space that are actively talking about AI in podcast and we can invite them to be part uh or to take part um in our uh in our podcast.

[00:23:28]That's going to be pretty much what's what we're going to do and I think the problem in this specific case is how do we get their attention? That's another problem that we going to talk about.

[00:23:39]The second thing is that we going to choose a platform. Now, that's going to be funny, but I think in this case, which I usually don't do, I will actually use Open Claw.

[00:23:49]And if we go back to the slide where I explain about type of AI agents, the reason I'm going to ask Open Claw to do that is because um I don't know what will work.

[00:24:02]So, I did not figure out it yet manually, so I can't come to an AI agent and say, "Here's a script on how to do that automatically." I am more in the type of a discovery phase, so I'll need to figure out I'll basically use AI to try to figure out how to solve the problem as opposed to automate it. Um we have a question from Rotem. Let's go with Rotem. Yes.

[00:24:23]>> Hi. Uh quick question. Sorry, there's two of us here. Why do you usually not use Open Claw?

[00:24:31]>> Sorry?

[00:24:32]>> You said that you usually don't use Open Claw. What's the reasoning behind usually not using it?

[00:24:37]>> As I mentioned, it's not about I I I'm just saying that in this type of of project I wouldn't use Open Claw, but as I mentioned again, there are two types of agents.

[00:24:47]Uh in in this case we chose a problem that I don't have a solution for, so I'll use Open Claw in this case because I want to find a solution for it.

[00:24:55]As opposed to building something that I have an assumption of knowing exactly how to solve this. So, the reason I'm going to use an autonomous agent as a as opposed to an agentic workflow is because an autonomous agent in this case will also help me find a solution for the problem as opposed to try to automate the solution that I have in my mind. So, that's basically the way to think about it. Does it make sense?

[00:25:18]>> Yeah, thank you.

[00:25:19]>> Any other questions before I'm diving in?

[00:25:24]Cool. Let's go.

[00:25:25]Um I do recommend a company called Claudy.

[00:25:32]They're great great great company. It's claudy.ai.

[00:25:35]They basically, and I'm going to open it up here. Um they basically are hosts.

[00:25:39]So, think about it as a you know, kind of a digital ocean from the old world that just like a host for Claudy and Hermes.

[00:25:47]We're going to talk about Hermes as well, by the way, in a second. Uh but the reason I'm using Claudy is just because they have an incredible onboarding. Uh they're great. They're based in Silicon Valley. Um very um uh responsive, by the way. So, if you have any bugs and stuff, um they're very much fun to work with. Um and we're going to use them to run an open claw on a server.

[00:26:08]Now, the other thing that I need to think about right now is So, now we have the platforms. Now, I need to think the tools.

[00:26:15]I think I actually need to think about where do I find um people who are actively interviewing for podcasts. I've never even thought about this before. And then, how do I feed my open claw with those tools so it can perform the task?

[00:26:36]Here's what I'm going to do. I'm going to go to Autonomous Capital.

[00:26:39]And the first thing I'm going to do is I'll say, "I'm trying to find um search, right?

[00:26:46]There are different search capabilities such as photo search and 64 and Tavily.

[00:26:51]Tavily, by the way, is an incredible company that just got acquired for half a billion dollars. They're today one of the largest infrastructure tool using the web. And I have an assumption here that I think most podcasts are on YouTube. And so for that reason, if I use Tavily, which is basically connected directly to the web, I will be basically able to find people who are actively interviewing for podcasts. And then the agent will be able to find their information. So, what I'll do right now is that I'll log in.

[00:27:25]And I think I have an account for my company, so that makes sense. We'll create a new API key, and you are not going to copy it because it's not cool.

[00:27:36]Okay, so we're going to go to the API.

[00:27:38]So, basically what Tavily does in this specific case, it just allows my OpenClaw to search online.

[00:27:45]Now, OpenClaw usually does have a connection, but this one really does well. When it really does well, it connects directly to Google. So, whenever I want to use YouTube, for example, I'll use something like Tavily. There's another one that is pretty good that is called SerpAPI. Also has different integrations and APIs, but it gives you the ability to, for example, to specifically say Google Videos API or YouTube API. And then if I only want to do YouTube and search on YouTube, I'll give my OpenClaw this. So, I'll go to here. I'll go here, sorry, and I'll say, "I need a new API key." We're going to call it Mickey test.

[00:28:26]And we're going to set it as development because I don't want to go crazy. Now, I have an API key. You can copy it. I'll delete it later.

[00:28:33]And basically, I'm going to open up my OpenClaw. So, this is my OpenClaw.

[00:28:38]There's nothing very interesting about it. You can, by the way, open it up in any possible way. So, you can go to Telegram, for example, and chat here. You can open it up on WhatsApp. You can open it up on email. It doesn't really matter. And your chat interface can do that. And Open Close, since since it's an autonomous agent, will basically help me to find a solution for this. The only thing I do want to make sure is that I'm connecting this um tool Tavily to Open Close. So, that's what what I'm going to say basically is I'm trying to find a way to get new um people to get to be interviewed for my podcast.

[00:29:22]Uh it's called Be Autonomous Business, and I host people for leaders in the AI space.

[00:29:33]And I'll basically say, "I want you to use um Tavily. So, I'll go to By the way, the way you actually install a tool like this is basically you go here, and you just copy the documentation.

[00:29:49]That's it. So, you copy the URL, and you tell Open Close, "Please use Tavily to um uh search for people online.

[00:30:04]Here's the doc, and I'll say, "Here's the API key." That's basically the only thing it needs.

[00:30:16]And another thing that I need to think about right now is how do I make sure that um Open Close has a way to send emails, right? Because that's the next thing.

[00:30:24]So, it's going to find people, right?

[00:30:26]But then it need needs to actually um reach out to them. So, I'll do right now I'll do this right now. I'll ask it for a list. Give me a list of people that might be relevant um New York or Tel Aviv is best as a location.

[00:30:50]And we'll see what happens next.

[00:30:53]So, we basically gave it all the instructions and we just realized that we think we know which um, uh, tool is best for this. It might take a little bit of time just because sometimes the host is a little bit overloaded.

[00:31:08]Um, and then it will use different tools, as you can see here, um, and try to come up with a plan to to achieve this.

[00:31:17]The next thing we need to think about in the meanwhile what it while it's working is how will my agent send out messages?

[00:31:24]Now, there are many ways to go, okay?

[00:31:26]So, for example, there is a tool called Unipile, which you can find also on autonomous capital website, um, which is a unified API for messaging, email, and scheduling. I usually use this for LinkedIn. So, if I want to send automated messages for my LinkedIn, I usually use Unipile. They're French company, uh, doing great, um, and they're great, great, great, great opportunity for you if you want to connect specifically your LinkedIn. I can also go and say it'll probably find emails of people.

[00:31:54]So, I want to make sure that my agent has, um, an email inbox. So, I will say, "Hey, I want an email inbox for my, um, AI agent." So, that's probably what I'll go with. And the same way I, um, installed before, um, Tavily, I'm going to install the agent mail thing. So, I'll go and I'll ask for the, uh, documentation, which is right here. It's always on the left side.

[00:32:21]And I'll say, "Here's the type of thing I need." Um, and I'll ask my Open Claw to connect to this. Let's see. It still takes time.

[00:32:30]Um, so we'll have to either wait or I'll fill it up with other stuff.

[00:32:35]There are off-the-shelf So, do it again, when you're thinking about agents, either building or buying, when you're thinking about buying, there are off-the-shelf agents that I highly recommend. One of them is actually called Manus, which some of you might know, which is an off-the-shelf pre-made agent that already has some of those tools. So, it knows how to develop desktop apps, build websites, create slides, or do research, and then I basically can do the same thing by buying the exact by buying the service specifically from Manus. Manus is was acquired by Facebook not a while ago.

[00:33:10]So, it basically can give you like off-the-shelf capabilities if you don't want to deal with OpenClaw and all these kind of stuff.

[00:33:18]The difference will be is that it will be harder here to connect external tools because they have a preference for internal tools that already were integrated by while it was built. And of course, as I mentioned before, the other way to go is to go to Lovable and say, "Hey, I want to build an app that will find people online who might be and you know what, I'll go to my OpenClaw and just do this.

[00:33:49]And this is going to be an AI agent as well. No actual difference between those. So, all of them are writing code right now trying to figure out how to do something, right? All of them.

[00:33:59]But all of them are trying to to do that in completely different ways. So, the hard part about agents is actually finding which platforms {slash} which tools are best to solve the specific problem that you have access to. And the reason this is extremely important is because most use cases, and we at Answer are building lots of GTM agents, are built off a completely different infrastructure.

[00:34:27]So, every use case has different tools.

[00:34:29]And the probably the way to think about it is that the more tools are being released to the public, the harder it is to actually build an agent because you actually have to think more about different type of tools and how to orchestrate them as opposed to actually going with one tool and just making sure. So, it did found it did find a a list of of um people that it thinks would be uh incredible for my um for my um uh podcast. And I think what we're going to do next is we will connect a an inbox to my OpenClaw and we'll actually start sending emails. But, the difference will be is that I'm not going to tell it how to do that. I'll just say that my KPI and we'll just start like this. Okay. So, we will go to API keys just like just like we did before. We'll create a new API and we'll create an API key.

[00:35:27]This is great.

[00:35:29]And then we'll say, "I want you to set up an inbox for yourself.

[00:35:41]Let me know what will be the the the address so you can reach out on my behalf to those people and pitch them about my podcast."

[00:35:58]API key, again, same thing.

[00:36:02]And then the second thing I'm going to do is I'll say, "Uh here's the documentation."

[00:36:10]So, what OpenClaw going to have right now is that we'll go through the documentation, try to figure out how to create a new inbox. We'll probably see it right here as an inbox on AgentMail. And that means that the agent will actually have an entire um inbox of itself. So, it can reach out on your behalf, it can do stuff, it can reply if they reply, it can send it my Calendly, whatever. It can basically have an entire conversation uh over email with its own inbox with those people. So, basically we just solved a big problem for uh myself um and which is basically meaning how do we get really fun and incredible people on my podcast. But, the difference will be and that's an autonomous agent and that will be the difference is that the next thing I'm going to say once it's going to have um its own inbox, I'm going to say, "Here's the KPI. I want you to get at least one person um um on my podcast until end of May, for example. So, send as many emails as you need, um iterate on as many copy as you want, and let me know what worked." So, people always forget that the main idea about agents is that they can learn. And this is an incredible opportunity for you, too. So, for example, if you want to solve a problem and you don't have enough time for this, an agent can help you find a way to make it work. So, in this case, instead of trying to figure out how to how to pitch people about my my podcast, I'm going to let the agent do that. So, I can put it on the side and I can get daily emails about, "Here's my progress. Here's how many people I reached out to.

[00:37:47]Um and here's the copy that made people uh reply." And this is an incredible opportunity that we as just human beings cannot actually figure out because it's usually extremely hard and we are uh tend to think about those stuff um by ourselves.

[00:38:02]Um amazing. It just created an inbox for itself. So, we're going to check it out.

[00:38:08]It says, "I'll use an dedicated uh um thing. Um that's my Sarah autonomous business at agentmail.to is my URL. You can, of course, go to agent mail um and refresh and see it. Probably going to see Sarah. Here's Sarah.

[00:38:26]And we uh you of course can use like different domains. So, for example, if you don't want it to send emails from um from this weird thing, uh you can ask to to have your own domain, which is actually what we're going to do cuz I have a verified domain. I'll say um I have a verified domain on agent mail.

[00:38:50]The two it's uh whatever. So, create a new one with this domain and send a test email to [email protected].

[00:39:09]So, what I basically did right now is that I created a list of people to reach out to with Tavily because Tavily has access to YouTube, so it had uh the chance to actually review a few podcasts and see who's actively interviewing for podcasts right now. And the second thing I'm doing is that I'm creating an an inbox for the agent so it can actually reach out on my behalf, and I connected my own URL um on um agent mail. So, basically what's going to happen next, sorry I can't find it. Basically what's going to happen next is that you will actually see you'll see it right now live that once it's going to create a new email inbox, it will actually send out emails directly to my email so we can test it out. Once we're going to do that, I'll give it the go and I'll say, "Listen, here's what I want to do."

[00:39:55]Okay, so Sarah autonomous business.

[00:39:57]We'll go to my specific email. We'll see that if we got it.

[00:40:01]Did we get anything from this?

[00:40:04]Why we did not? Let's see.

[00:40:08]Maybe it's spam.

[00:40:09]It might be because I've been playing with this. Here's here here it is. Okay, it looks safe.

[00:40:14]Um so, we got an email from Sarah. Sarah is basically the agent that we just created a second ago. It goes to spam because my URL that I used here has been abused a hundred times.

[00:40:25]Um and that's the way it works. But, if you open a new URL, it's not going to go to spam and we're going to go after this.

[00:40:32]So, basically we have everything set up.

[00:40:34]We'll say, "Oh, and another thing is that you do have a UI right now on Agent Mail." Uh which I always forget the thing of. Okay, now we have a new thing.

[00:40:45]You can actually go to sent, so it looks like a an inbox and you can see what Sarah sent over. So, any uh reply or anything that will come into this inbox, you can basically view as a UI here.

[00:40:57]But, the agent doesn't need a UI. They have it as an infrastructure. So, to close the loop, I'll just basically go and say, "Hey, okay, perfect.

[00:41:06]Um I want you to find the best copy. That's the difference, by the way, between autonomous agents and not, as I mentioned, that will convert those people into hosts uh into guests on my podcast. You can read about my podcast here.

[00:41:35]We'll go to say one of the stuff that I did.

[00:41:42]Brian King from A16Z. We had a great great great interview together. Perfect.

[00:41:50]So, here it is. This is me.

[00:41:52]Um we're going to go back to Open Claw and we'll say, "Here's where you can learn from. Please let me know once a day how is this going as a report to [email protected]. That's it.

[00:42:14]Okay? We actually have an agent that will right now try to figure out how to get this list of incredible people, which I don't really have the chance to review, but incredible people, Pine Cone, heavy uh runway, one of the best uh some of the best AI leaders in the world. But, instead of letting it know how to do that, I'm using AI to solve a problem by design. And that's a completely different approach as opposed to saying, "Go to X, scrape Y. Here's the copy that works for me." This is a completely different approach, and that's why I chose Open Cloze as opposed to um you know, N8N, for example, where I define the specific um um steps. So, we just had a pretty much live demo of how you can utilize an AI agent, and but also the type of tools that that you can use, but I think the most important part here is what does it mean to think about designing an agent?

[00:43:16]Because agent by the end of of the day is something that is has no UI most of the stuff most of the in most of the cases. You can build a UI if you're using live coding platforms, but it basically is designed to solve a specific problem, and that is basically what I'm focused on. Um I have a bunch of questions here I'll try to address.

[00:43:36]Uh wow, okay. Open Cloze is very unstable from my experience. Well, all of this technology is very unstable from my experience, and that's the cool part about it. Um you can experiment stuff that other people did not have a chance to do, and so for that reason uh um I'm highly I'm I really recommend using it. So, um this is just [clears throat] important.

[00:44:00]Um pref- my preference is is Hermes. Hermes is great because it has the self-healing component. That means that it basically tries to solve itself uh solve problems by itself. It's also cool because it has an a very cool technology for a very long context. So, for example, if you have um you know, a a Reddit thread, for example, I'm writing a lot about this and you want to eat all of it. So, Hermes is much better for this, but Hermes is much slower. So, it is something that I would take into consideration. Sometimes it's just extremely annoying. Uh switch to Opus? No, it's expensive. I'm trying not to. Uh Opus 4.8 Yeah, 4.8 was released like an hour ago, so I didn't have a chance to play with. How do you know what tools you're looking for?

[00:44:49]So, you can ask the agent and you can just browse the web for, you know, but you basically most cases um trying to use Think about it as if you would do the job, what type of of tools you would have would have to have a an email cuz you want to send emails.

[00:45:08]That's one. If you want to search the web, two. Think about you doing the work and then try to find the tools um for agents. OpenClaw remembers all these preferences from one session to the other. The answer is yes. It has a specific component um that um uh kind of overrides itself, but in most cases when you're using a hosting, they have it turned on. So, if you're using it on a computer, it would probably be uh harder. Um can you talk about the risk of OpenClaw? Yeah, the number one risk of OpenClaw is that it can be hacked because if you go to my OpenClaw and I expose it to the world and you ask it for information and you manipulate it hard enough, you can get all of my information um out.

[00:45:58]The way I'm thinking about it is I rather risk myself with exposure than not use the most advanced technologies because I'm losing more than I'm winning.

[00:46:08]Okay?

[00:46:10]Um so, that's usually what I'm thinking about it.

[00:46:12]Um for podcast booking, is there a moment when agent hands off to you? Yes, it's called human in the loop and it's your decision when it happens.

[00:46:22]I usually when I do stuff like this, I just give it a Calendly and I say, "Whenever you hear a positive sentiment from the person and if he's interested, um just send over the calendar." Usually people like it.

[00:46:34]Um that's great. Budget-wise with all of these tools, adding more tools and more monthly subscriptions, you don't know what is the budget will look. The answer is that's true and we don't even talk about the tokens. The tokens are even more expensive. So, what happens is that you have to think about the value um in the same way. So, the best way I have in my company, we have something that I called the Upwork test.

[00:46:57]If you would hire somebody who's not an AI on Upwork and pay them X, what would it be? And if the answer is $3,000, then don't go over the $3,000 with tokens and and tools. If the answer is $10,000, the same. If the answer is $300, the same.

[00:47:12]So, I highly recommend you utilizing the Upwork test because I think it's a great way to think about it.

[00:47:21]Um Yes, we have experienced creating integrations into social media without getting blocked. I will talk about it in just 30 seconds.

[00:47:33]What people are usually doing is that they're trying to spam any social media and no social media on planet Earth likes it.

[00:47:40]Um so, that's why you're getting banned.

[00:47:42]The way to think about social media is that, you know, every social media wants engagement. So, if you are contributing to other people with your agent. So, I have an incredible agent on Reddit. It does extremely well and people like it.

[00:47:55]Uh you know, they upvoted, they comment on it, they ask questions, they have the entire thing.

[00:48:00]Um so, um the way to go about this is don't be too promotional. Don't use the direct API for this because that's a very bad way to get banned. And uh I do have a few more tips. So, whoever asked the question about Reddit specifically free feel free to reach out. I'm happy to to guide you through. I will say that we as a company um you know, we're a growth hacking agency growth hacking lab and a lot of those researches about how to use how to go into Reddit, how to go into Wikipedia, how to work with social media, how to work with SEO and GO, we just have as research open on our website. So, you can go to answer.bot, which is right here, and just, you know, read it. We have everything open. That's the idea of the company.

[00:48:43]When will you transfer Open Claw POC?

[00:48:47]We uh I don't see Open Claw as a POC. We have customers using or we are using um Open Claw for customers.

[00:48:55]It just you have to know how to do that.

[00:48:57]There's a specific environment, um specific type of infrastructure, specific type of um workflows that you would and wouldn't do, data.

[00:49:07]Um so, I don't see Open Claw, especially if it's hosted on like a virtual machine, as a non-production thing. It's good enough to solve many of the GTM cases of large enterprise customers for us today. So, I I'm I'm big into this. I'll answer another question then I feel free for you to just jump on um on on the on the voice and ask questions directly because um it's more fun.

[00:49:32]What's the autonomous site that lists?

[00:49:35]Yes, Autonomous Capital. It's um NFX plus myself um a venture. We're investing in agency tools, but you can find agency tools on the site just as a resource.

[00:49:48]Um yes, that's it so far. Any other questions, Feel free to um hop on the stage and ask anything you might have and I'm happy to make this as a as as as a conversation. Yes, David.

[00:50:01]>> Hi. Uh thanks so for your presentation.

[00:50:04]Uh I have a question on when would you consider using fine-tuned models for specific tasks like what what are use cases from your personal experience or some like company use cases when you actually see like huge value in fine-tuning local models?

[00:50:23]>> Um the answer is there are two ways to like there are there are two metrics to look at. Number one is speed. The whole fun about fine-tuning is is that it can do things faster because it's usually guardrailed with different you know, limits. Um so, for example, if you're working on a um medical thing and you need to answer extremely fast because it's a doctor that now queries the LLM and they need to um save somebody's life. That's a a good use case to look at. Second thing is a very curated specific type of content.

[00:51:04]So, I know I've been looking into a cool project that does uh DNA um also fish medical thing.

[00:51:15]Um so, to query a DNA it's extremely hard because it has hundreds of billions of different um um data So, you don't want a general LLM. So, you know, knowing what is the capital of Paris when you're looking at a DNA, but you do really want to make sure that a DNA is encoded into that into it. So, I think it's something that will be as you know, very very very much um um specific of a of a very very much deep use case. That's basically the way to go.

[00:51:44]Other questions?

[00:51:49]Okay. people are asking for discovery of models.

[00:51:53]Open router is a company that lets you optimize price for uh um for like LLMs.

[00:52:02]Uh I'm going to share my screen.

[00:52:04]So, here's Open Router. Um it basically, you know, a proxy for APIs that basically allows you to switch between LLMs quickly. If they're very expensive, so for example, if you're using Altus and you're just running out of money and you want to make sure that whenever you're running out of money, use a DeepSeek, uh that's the way to go. But they have a page that is called rankings, which I highly highly highly recommend because it just has rankings of different APIs that Sorry, LLMs. And you can find like the weirdest LLMs in the world, like Cheemee, which I really like. It's like a Chinese niche thing.

[00:52:40]Or or like um Owl, which is like a very specific LLM that is super cheap and fast. And And experiment experiment with all of them.

[00:52:51]I actually have one use case that people don't think about, but it is important.

[00:52:55]So, for example, you want an LLM to write or talk specifically like you.

[00:53:01]Um I usually am not using, you know, Claude or OpenAI because they have many rules and and and exceptions and specific design. And it's not something that is happening with DeepSeek. DeepSeek is more flexible, so you can basically turn it into whatever you want. So, if you want it to speak in a specific way or to behave in a specific way, it just has less barriers.

[00:53:23]The less popular the the LLM is. So, just go crazy. There are thousands of them all over the place.

[00:53:30]Um yes. I appreciate your time, guys. We're running out of time in just a minute and a half. So, one more question and then we'll wrap it up.

[00:53:46]>> No questions.

[00:53:49]Mike wants also to look at artificial analysis. Yeah, yeah, that's good as well. It's same type of of of the solution.

[00:53:58]This is my email.

[00:54:00]It's right here. I am doing this literally all day long, 19 hours a day, and I'm a fanatic.

[00:54:08]And and crazy person who does AI all day long. So if you have any weird challenges that you're experiencing with AI or you're just thinking about the tools that you need to use to solve a problem or which platform, I'll definitely I'm definitely happy to answer. So any questions, feel free to reach out. Pete, I appreciate the opportunity to pitch and to share my knowledge with the NFX community. I appreciate it. Thank you so much.

[00:54:34]>> Thank you so much, Miki. Awesome.

[00:54:37]>> Enjoy it. Have a good one.

[00:54:38]>> Yeah. Have a good day, everyone.

[00:54:40]>> Bye-bye.