Abhishek Das of Yutori on Building AI Agents to Take Over Your Digital To-Do List

Imagine a world where AI agents handle all your tedious online tasks – from tracking flight prices to securing restaurant reservations. That's the vision Abhishek Das and his co-founders at Yutori are building towards. In this episode, Abhishek explains how Yutori's AI agents, called 'Scouts', monitor the web on your behalf, freeing you from the endless cycle of refreshing tabs.

He dives deep into the technology behind these agents, revealing how they navigate the complexities of the web using a combination of APIs, search tools, and a proprietary in-house navigator agent. Learn how Yutori is tackling the challenge of building truly autonomous AI agents and transforming the future of web interaction.

Nataraj (00:01.509)

Hello everyone, welcome to Startup Project. Today we are joined by Abhishek, he's the co-founder of Utori. Utori has really high pedigree co-founders, Abhishek, Devi, Nthro, all being ex-meta AI division. I think Abhishek himself has a PhD from Georgia Tech. And Utori has been building autonomous AI agents that can handle our everyday daily tasks. The first product scouts by Utori.

The plus a team of AI agents that track anything that you want on the web from flight prices, restaurant reservations, sales leads, job interviews, job openings, and basically help you save time in terms of browsing the web. There is $15 million from an interesting group of investors, including K.F. Lee and Jeff Dean, are the icons of the AI industry. This conversation will explore the opportunity

to go deep in terms of what we can build with agent-ic AI, what are the technical challenges behind building useful AI agents, and how is this technology going to transform how we interact with software, and more importantly, I think the web. With that, Abhishek, welcome to the show.

Abhishek Das (01:21.006)

Thanks for having me.

Nataraj (01:22.501)

So give us a little bit background, you how did you totally got started? How did this team of X Meta folks farm and from this company?

Abhishek Das (01:33.398)

Yeah. So I just as background, I grew up in India. I moved to the US in 2016 for my PhD that was at Georgia Tech. That's where I met my co-founders like now co-founders Dhruv Bhatra, who's one of my co-founders at UDOT. He was my PhD advisor at Georgia Tech. and Dhruv, they were both faculty at Georgia Tech. So the three of us worked together a whole bunch as part of my PhD.

At the time, so this was in 2016, 17 as part of my PhD, I was working on problems at the intersection of computer vision, language, and reinforcement learning. So if you think of a home robot and asking it the question, where did I leave my keys? In order to answer that question, the robot would need to understand language. It would need to know how to navigate an indoor environment and then come back to you with the answer or with the object that you asked about. And at the time, I was looking at, OK, how do we train deep learning models for these kinds of tasks?

Again, is back in 2016, 17. So AI systems weren't anywhere close to as capable as they are today. It was a fairly exploratory space to be working in. Like the first set of sort of deep learning image classification models around like AlexNet, VGG, et cetera, those had happened. And then 2015, 2016 was around the time when image captioning was a pretty big deal. So generating a one sentence description about an image.

So that was sort of state of the art at the time. And I was kind of pushing it towards being more embodied that, like if this agent not only does it see and talk, but what would it take to make these agents act in the physical world? And initially I was of course like trying to develop these methods entirely in simulation, but I did collaborate on some projects where you put these on actual physical robots as well.

I had a lot of fun as part of my PhD. spent time at like on internships at a bunch of places like DeepMind, FAIR, Tesla Autopilot and so on. Worked with a phenomenal set of people. I finished my PhD in 2020. After that, I joined FAIR at Metta as a research scientist. Yeah, I worked there for like about three and a half years till early 2024 when David, Dhruv and I, all three of us decided to leave Metta to start Yadodi.

Abhishek Das (03:55.086)

Davian's Roo, they've actually had joint positions as faculty at Georgia Tech, but as well as research scientists at Metta since 2016. Yeah, so like we've known each other for almost 10 years at this point. I think since we started having a weekly sort of brainstorming session back in like 2017 or 18, when we would just meet once a week.

to discuss that, okay, like if we were to start a company at some point, what would it be about? What would we go after and so on? So it's not like we've been just brainstorming for like 10 years, seven, eight years, like over time, it just became a casual social hang. But that's sort of how early we had started thinking about starting a company potentially together.

Nataraj (04:44.869)

So how did the idea for like, scouts came about?

Abhishek Das (04:50.434)

Yeah, so with Yutori, when we left Meta, it wasn't the case that we had a clear-cut idea of what we were going after. The plan was that, OK, let's start prototyping, brainstorming, building a few things, see what sticks, and take it from there. And that's basically what we did. So we left Meta, Devi and I left Meta April 2024. And we started writing up a few ideas, discussing it. We built a few prototypes. And very early on, we had written down this idea of a

of an AI concierge that could handle a few digital errands or a few chores on your behalf. And we had kind of moved on from it to other ideas. But we kept coming back to that one from different angles. An additional bit of context here that might be useful is that all three of us in some capacities care quite a bit about our very intellectual productivity tools.

And so like anything that can take mundane work off our plates or like any tools that helps increase our productivity. Yeah, we've been sort of experimenting with different kinds of tools and methods over the years. Like Devi has a blog post on a very opinionated blog post on how one should use their calendar that went viral and she had written it up I think several years back.

I had built a very simple sort of note taking app, which was also like pretty opinionated on like the sort of affordances it offers versus not. It's almost 10 years old at this point. Yeah, so we had, to varying degrees, we've delved into like productivity tools. So that intersecting with our background in agents sort of eventually evolved into what we are going after with Utor.

So with Utori, what we're building are basically, like you said, AI agents that can do everyday digital tasks on the web. So anything that we currently open a web browser for that one would broadly categorize as digital labor. So one category of things that we open a browser for are entertainment, right? To watch a show on Netflix, let's say, right? But this is not that. This is if you are trying to get a task done on the web, it feels like...

Abhishek Das (07:11.434)

In the future, five to 10 years from now, how we get that done is going to look very different from what it looks like today. We may no longer be manually clicking buttons, navigating web pages, scrolling on web pages, have like 100 tabs open. We may be operating at a slightly higher level of abstraction, where we are talking to AI agents. And agents are the primary drivers of action on the web. And so our not-star is to push on that frontier and make that

vision sort of a reality. And we do this, sorry, go ahead.

Nataraj (07:44.198)

So in 2019 or so, I wrote this blog post called the Push vs Pull Model, your brain is hijacked by push notifications. Everything is pushed to you, which song you want to listen to, Spotify recommends to you, YouTube will recommend to you which thing to watch next, and you're browsing for anything. Everything is pushed to you. So you're constantly juggling with

content that is being pushed to you and people and that's why like you see like the choices people make are actually being reduced in some form like personalities are also becoming if you have like thousands of different personalities as human beings now we have hundreds of them because everyone is sort of like mimicking each other because they're fed by same content right and that's why you also see increasing in

polarization also because like it's easy to put in three buckets as a person rather than in 300 or 400 because people used to consume things differently based on their curiosity but now people consume based on what is being pushed to you. So my whole point was like how do you have a pull system? How do you like go from you know someone else telling you or is it like you follow your intuition, your curiosity but the tooling was only there for if you can build things if you can like scripts or it if you can

know, hijack, you know, or put together a bunch of tools together. If you're a software developer, you could do that. It's also like there's an element of, can you pay for stuff? Can you avoid the ads? Can you avoid the recommendation systems? Like, can you take YouTube subscription where you don't see any ads? So there's almost a certain segment of population who has evolved above that thing where they will never see an ad because they have all the subscriptions, right? So...

Right now what happens is I don't have to be on Twitter to know what is happening. I can just write an agent that is going and telling me, find me top 10 important AI tweets that I want to know from XYZ researchers. And I can have a scout or I can have deploy another type of agent. can write a Python code. I can just code this up and ask Repletor to do something. So I think the interesting thing...

Nataraj (10:06.935)

The point I'm really trying to make is you go from that someone else telling you to what to consume because you're in that web interaction system to you're going towards where you can, you have more control about what you want to consume and what do you want to stay on top of. And I think Scouts really helps us do that. So I think before we jump ahead of the conversation, I think it would be good to sort of like give a small demo of how Scouts work and then.

We can continue the discussion from there.

Abhishek Das (10:38.126)

Sounds great. And the push-hush versus pull model, yeah, definitely resonates.

Abhishek Das (10:47.916)

Okay.

So hopefully you can see my screen. This is what the home page of Scouts looks like. As context, Scouts is basically agents that can monitor the web for anything you care about so that you don't have to sit there refreshing tabs and monitoring various things. You set up agents, they'll monitor and they'll notify you whenever they find something relevant.

Nataraj (10:52.835)

Yes.

Abhishek Das (11:17.294)

So I'm just gonna go ahead and set up a few scouts. let's say.

Abhishek Das (11:25.418)

Any time someone mentions or reviews Scouts by Utility on Twitter, LinkedIn, Reddit, etc.

So I type in a query, it suggests a few improvements. We can skip that, that's optional.

Nataraj (11:45.167)

That's actually one of the most useful things I've found while using because you frame something and then realize, this is actually better framing. So it's actually a very good product feature.

Abhishek Das (11:56.377)

Glad to hear that. And then the second thing it asks for is, how frequently do you want to be notified about it? Do you want a daily digest, a weekly digest, or whenever there's a new mention or review of scouts? So for now, I'll just pick whenever. I'm actually going to go ahead and set up a few more scouts, just so we have all of those running.

So let's say whenever there's a new browser or computer use model, especially open source releases.

Abhishek Das (12:31.406)

That's another thing that's top of mind. So I set that up. It can also do a good job of monitoring product prices or reservations. So let's say a lunch slot for two at this restaurant called Gopro in SF, which is an Indian restaurant that I really like for Saturday or Sunday.

Abhishek Das (13:02.572)

And there's a bunch of examples here just to sort of solve that cold start problem that gives you ideas for what to scout for. And there's a bunch of existing scouts that we have kind of curated that give you a flavor for what the product can do. So once we have these scouts set up, that's it. You don't have to do anything else manually. It's basically set up. It's on. One thing the UI lets you do is you can see.

behind the scenes what it's doing to generate sort of the first report for that. So here, this was my original query. Anytime someone mentions or reviews scouts by utility, you can look at, OK, there's a few subagents. There's a social media subagent that's looking at right now Twitter, YouTube, Reddit. There's another subagent here that's doing a Google search to look for mentions of scouts. So it's going to do its thing. It takes a little bit of time.

Whenever it's done, it's going to send a report over email.

Abhishek Das (14:08.366)

And just to give you guys a little bit Sorry, good.

Nataraj (14:08.441)

I think when I started using the Scouts, there was no this UI. I think all it was doing was it was just deployed and you see progress or it's done and you get an email when it's done. But this view actually is very interesting because we also get that report of like how many agents were deployed, how much time it was saved. And that was also very, really useful in terms of like quantifying how good of a Scout you deployed.

or not because like, is it worth my time to have this come or not? Because that time saving aspect of it was also an interesting feature.

Abhishek Das (14:43.788)

Yeah, exactly. And this live stream UI we shipped recently. So it makes sense that the last time you used it, you may not have seen it.

Nataraj (14:52.037)

So behind the scenes, let's talk a little bit about what's happening behind the scenes before the scout does a job, right? So how does the technology work behind the scenes? Do you have your own version of the web or you're just plainly deploying virtual agents in the cloud that are going and literally browsing it? What is the type of approaches that you're taking here?

Abhishek Das (15:15.446)

Yeah, exactly. there's a few things, a few different things that are happening under the hood. And this view kind of already gives you a glimpse into it. But there is more than 100 MCP tools that the agents are using under the hood. Many of these are happening in parallel. Like you can see here, there's three subagents working in parallel. And for the long tail of the web, where so the long tail of websites that don't have APIs or MCPs,

where you need an agent or a person to actually browse or navigate that website and click through pages to get to the right piece of information, we have our own in-house navigator agent, which is basically doing that. So if I skip to the back here and look at what it's doing, it's clicking through various web pages to get to the right piece of information so that it can report it back to me.

All of this agent orchestration is built in-house. The navigator model that powers this subagent is also built in-house. We are in fact currently state of the art on several browser use benchmarks compared to like Anthropic, Gemini, et cetera. But that's basically how it works under the hood.

Nataraj (16:28.709)

Yeah, was trying to, don't know if you know, Praag Agarwal is the ex-Qriir CEO and he founded a company called Padlan. And I was having a conversation with him at an event in Seattle and he was mentioning that they basically recreated the search indexes, sort of like what Google has or what Bing has for deploying their agents or for their agents to work on and.

Abhishek Das (16:37.519)

Yep.

Nataraj (16:56.217)

they continuously update their indexes so that they always have a snapshot of the web. So was curious if you guys are having a similar approach or is it more, is it not the direction that you're going?

Abhishek Das (17:07.907)

Yeah, so a couple of thoughts on that. One is that for monitoring specifically where you're looking for new information, an index is less useful.

For retrieving historical data, like answers to questions about historical data, an index is super, super useful because it can make things way more efficient and you don't need to solve it from first principles every single time. But for monitoring, it's less useful. And second is that that's not necessarily where we are looking to differentiate on. So like I said, this uses a bunch of MCPs and APIs under the hood. As you can see here, it's doing a Google, it's doing a call to Google Search API.

We use a bunch of other search APIs as well. So we're happy to use whichever APIs return the best results. The important bit here is to make this agentic so that it retrieves a set of results, it reasons about it, decides if it found the relevant piece of information or not, and then decided, okay, what to do next? And this is happening in sort of a for loop till it gets to the right piece of information, including actually navigating websites in case search APIs are insufficient for certain queries.

Nataraj (18:20.261)

So if, it is like a scout can be run per day or per week, right? Or like whenever a change happens, like Navigator is one way you're doing that. And then you're, this is researcher. What does the researcher do and how is it different from Navigator? Navigator is like a little browser instance where you're going and looking at it just like a human would. How is the researcher different?

Abhishek Das (18:39.279)

Yeah, these are just a few different subagents that we've kind of set up. yeah, the researcher subagent, for example, is calling a bunch of different search APIs to see if it can find relevant information about the particular query. You'll see, yeah, depending on the query, you may see more. I think we saw the social media manager subagent as well.

Yeah, so that's scoped to look at social media websites in particular. So each subagent here has a very well-scoped, narrow set of tasks and websites and parts of the web that it needs to look at, including a mix of APIs or navigators.

Nataraj (19:29.825)

I think let's go into one of those cards that you have that already worked and that we can just...

Abhishek Das (19:34.64)

Yeah, this is another account where I have a bunch of scouts set up. We can look at what the findings view looks like. So the findings view is where all the reports from all my scouts are aggregated in one place. So here's an example report. So this one is for top rated PS5 games under $20. It found that, OK, there's currently a discount running on this game, Saints Row 4. There's links here that we can click on and look at the listing.

Here's another example. So this scout was looking for any new news about health care and AI. So it found a couple of reports about this new announcement from Samsung India.

Abhishek Das (20:21.083)

So this is basically what it's doing. You can specify a bunch of different signals that you care about, and it will scour the web to find relevant information for that particular query and generate reports that look like this.

Nataraj (20:37.583)

Got it. And when you're running it on a daily basis or on a weekly basis, does the agent repeat what it has done? Like, let's say the first time I ask a question, it does it for the first time, right? Is it replicating the same thing or is it basically doing it from first principles again?

Abhishek Das (20:58.383)

It relies, it has some memory of the work it did previously. So the work it did previously helps inform subsequent runs. And that's useful to make sure that the user is not getting a duplicate or stale piece of information.

And it also helps inform which websites may have the right piece of information. like, for example, for the for the lunch slot query, if I'm asking about a particular restaurant next time, it doesn't need to start from scratch to find that restaurant's website. It can shortcut that to some extent to start on the on the restaurant reservation page. Yeah.

Nataraj (21:40.294)

I think that's a pretty good demo. I've been playing around with it and I've tested 12 to 20 scouts. I can honestly say it saved me a lot of time and it also allowed me to focus on things because what happens is this is important, but you search for it, you look at it, then you forget it. Then after 10 days, you remember that, this was an important thing I need to follow up.

and I need to again go to that website and see if anything changed. And that could be like, you know, looking at a redfin you're buying a house or if you want to sell your house and you're looking at like, okay, what are the sales that are happening around your house? Right? That you can just set up a scout to say, hey, look at all the sales that are happening in your county. And I think the other scout I was trying is I have a Microsoft stack and

terms of work, so I wanted to get productivity tips on a weekly basis and send me an email. So every Monday morning I get a productivity email about how to use the latest tools to see. And then I can see and pick up what is useful and what is not. Yeah, thanks for that demo.

Abhishek Das (22:56.515)

Makes sense? Yep. A couple of the scouts that we had just set up got back with reports. So we'll quickly show that. This is the Copra one. It got back with a few different slots that are available. There's a link here that we can click on to go ahead with the booking. And the open source releases.

Nataraj (23:14.885)

That is super cool. The one obvious next step that I thought, and I wanted you to let me know which product direction you are going is, but, okay, I have this Scouts thing, but I also want to move that information into something else or do more with it. Like, for example, okay, I have, let's say, I'm trying to grow on Twitter, for example. I got top 10 viral tweets in AI.

All right, my next step is going to be either I want to recode it or reply to it. So there's a lot of action to it or just I want to do a further summary of it and put it in my newsletter. Right, like there are now more things or in some ways you could say Scouts is mostly now reading and now you took that information, you synthesized it and then you want to use that information to do some action.

And that doesn't necessarily have to be in control of utoria or scouts, but something that I can control So do you guys have any plans to sort of extend the product in that direction?

Abhishek Das (24:24.931)

Yeah, so some version of that is coming. We are going to provide integrations for either looking at your existing context from your other apps or exporting Scout reports to your other apps where this might be one step in a multi-step workflow or some longer task you're looking to get done. And there's going to be two main surfaces where we provide that. We provide that in the Scouts product in some form, some

Yeah, we provide that in the Scouts product in some form, but we also have an API, which is optimal for developers to have maximum flexibility of how they want to, however they want to use the results of these Scouts.

Nataraj (25:08.077)

And as I'm looking at the subagents that you like once that subagents started showing up, also feel like the subagents are itself a product in some form. Right? Like for developers who are trying to do like different things, like it's almost like the subagent itself can be a product.

Abhishek Das (25:23.661)

Yeah. And we get that ask quite often from developers, especially where like they know for certain kinds of queries, they don't want the full breadth of subagents. There's like, they want a specific kind of subagent. how we are like, it is on our minds to like, work with them, figure out, what is the best way to offer that flexibility and control to configure various workflows while still using our sort of agent orchestration.

Nataraj (25:53.19)

Because at some point it also becomes like once you scale and once more and more users started using, like it also becomes like, okay, what is the cost of the tasks valued to me as a user versus what is the cost that is you guys are incurring? Are you guys spending more time on more agents when I actually don't want that much? And that would be an interesting dynamic that will also play out. where like, you know, if I know like you have like a fixed amount of subscription fee per month, but

I could give a really complicated scout and exhaust my $50 in February 2 scouts for you as a business. That would be completely negative.

Abhishek Das (26:33.411)

Yeah. And exactly. I think that's spot on. We actually have a blog about our agent orchestration and what that looks like and why it looks the way it does, which touches on this point exactly. one of the main reasons that we had to break this up, or there were multiple reasons why we had to have this sub-agent architecture that PowerScouts, but one of them was also cost. That if you have a single agent that has all the context for everything that it's doing,

that's expensive on a poor token basis at every step. Whereas if you just split it into different subagents where each subagent now has a very specific role and much smaller context, then just in terms of raw number of tokens, you might be spending it's way cheaper. It also keeps the context clean. There's way less irrelevant information in the agent's context. So the LLM gets confused way less. So there's many reasons for why that architecture makes sense. But cost is one of

Nataraj (27:33.392)

Have you, I mean, obviously you guys are thinking about agents much more deep on a day-to-day basis. Like how does the future look? Because we started with, you know, more of, I call it like agent workflows where, you know, the Zapiers of the world. And I use heavily an app called Relay that we feature on a podcast. Like Relay is amazing because it's sort of like Zapier, but on steroids.

like once you add LLMs on top of it and make the interactions and integrations really good. the product, basically it's more about how they design the product than just pure capability. So that's why I use it a lot for, and so most of the productivity I see is coming right now for most people is that workflow, agentic workflow level. And deep research is obviously really good. That's one type of agent you can think of.

Abhishek Das (28:23.439)

Mm-hmm.

Nataraj (28:28.197)

And then there's like the browser agents, which is sort of like the research I'm assuming is doing behind the scenes, some kind of browser agents. What are the other form factors that we can expect to see or have ideas about that might come?

Abhishek Das (28:42.191)

Yeah, I think one vector where we are pushing on, and Scouts is like a baby step in that direction right now, but there's going to be more coming, is proactivity in particular. So most of these agent products and chatgbt, Claude, et cetera, all included, are largely reactive. Like you come in, you enter a query, and then they do something.

And part of the reason why they've been reactive so far is because of reliability concerns. Like if these agents aren't accurate enough, reliable enough, then asking the user to give up control is just too big and risky of an ask. The user would want to be in the loop if these agents aren't good enough yet. But we are kind of getting to the point where they can do more. They can...

oversee your context proactively and look out for things on your behalf that you may not have the bandwidth or interest for. And they can do a pretty good job for many such tasks. I think what I'm looking, yeah, from us, from others in the ecosystem, there's gonna be much more in the vein of proactivity that's yet to come.

Nataraj (29:56.048)

Yeah, I remember I think Arvind from Glean has also been mentioning the product or they would be the next thing probably in the next couple of years. But I feel like people who are doing at the cutting edge developing products versus like people who are using it have slightly different opinion. Because I think I am a little bit on a cutting edge in terms of because I try a lot of new products. But I think people generally are not using that many of the agent tools. I think they're not crossed like the pre-search.

And like if I look at like enterprise adoption or even just a regular consumer level adoption, why do you think that is?

Abhishek Das (30:33.933)

I think on the consumer front, just mindshare is hard. even today, if you go just in the US, for example, ignore the costs and go in middle America and like ask which AI tools people are using, at best, chatgp is probably the only one that's on people's minds, right? Consumer mindshare is just hard. It takes a while for that stickiness to that sort of circulation.

to happen. And yeah, that's probably the main reason. A part of it may also, a secondary, important reason might also be just the ease of using some of these products. if, yeah, like it's one thing to have a universal interface where you can come in, say anything you want and get a response back, which is like a very easy to use an interface. But if it's like a somewhat complex agent builder,

or workflow creation like to, that's just too hard for a large set of consumers, right? So part of the challenge here becomes, okay, how can we offer powerful technology, but through very simple interfaces? Like for example, if I could shoot a text to an agent and tell it that like my grandmother's out at the grocery market shopping, check with her whenever she's done and call an Uber for her so that she can come back home.

That's a very simple ask that's not possible to do today. Because that kind of multi-app orchestration, there's some parts that may be automatable but not fully. And the interface being as simple as shooting a text. It is solvable, but it doesn't exist today.

Nataraj (32:23.545)

Yeah, I think competing in consumers, you know, it's a tough beast in some sense. mean, B2B has its own challenges, but I think it's easier to convince of value and mind share wise, because you can convince a couple of people in the company and then they will adopt you and try and give you some audience. But I think consumer is especially challenging. that's why like, even perplexity in somehow, I think has this problem of like, not crossing the chasm of early users.

In the coastal areas, people are tech savvy, they use perplexity, but Chagy Bidi is directly competing with them. And a normal consumer who's not tech savvy cannot see the difference on why they have to switch from Chagy Bidi to perplexity. And especially when the whole memory layer is now sort of becoming the moat. Because Chagy Bidi knows me, Chagy Bidi, there's so many interactions. And once the memory is enabled, it's easier to just ask.

Like I'm filling out a questionnaire, let's say, and they'll ask me, know, hey, give me why you're good at particular this. And I'll tell Chargivity, use these experiences that I have and like to put out, because if I'm doing it on a new app, I have to give them that context of what projects have worked. While Chargivity knows all the products I've worked on, it has a couple of documents or even my resume that I already worked on, which has the details and that.

memory is so important and I think that actually is the real mode in consumer right now.

Abhishek Das (33:55.332)

Yeah. Yeah. For us, like a lot of consumers who are, who come from non-tech backgrounds, either we've been, yeah, like somehow they've been able to discover our product or we've been able to reach them. They love the simplicity of it. Like they do like how simple it is to set up while still being quite powerful under the hood. And like, once they start getting their first, second, third report.

it kind of clicks that, OK, so this is how to best use the product, or this is what the product does. And that kind of starts clicking for them. We're seeing a small but decent amount of interest from enterprise customers as well, especially sales and marketing teams, who basically want to their top of the funnel work basically looks like monitoring for various signals happening around.

to decide who to do outreach to. Like it can be as simple as, okay, like if a company is announcing a series A, it's probably indicative that like their post-BMF, they're looking to scale, might be a good time to reach out to them. And setting up a scout seems to sort of fit in very nicely for that particular use case. So there have been a few people who've kind of organically discovered us, thought this was a great fit for that use case, told their team about it, and then...

we ended up signing on and onboarding them as like enterprise contracts. But yeah, like that's kind of what the spectrum of users looks like for us.

Nataraj (35:29.957)

Yeah, that's probably what I also figured out is when like if I want to monitor competition, if I also monitor for opportunities, right? Like even for your own career that you want to do different things. Let's say you want to write for a website or you want to be a contributing author to someone or you want to go on a podcast and see who are putting out your podcast guest opportunities, right? If you are the PR person who's looking out for you guys, then I would basically set up a scout.

looking for, okay, which podcasts are actually looking for, you know, CDC or, you know, C plus companies that they want to have and just have that run every day and look at the opportunities and then I can reach them on. Pretty much you can do this for, you know, if I'm a startup raising funds and I want to know which investors are actively looking. So on a regular basis, because all the databases will have lot of stale data, even though like there are bulk databases. I want to know what actively investing right now in the last couple of months.

then scouts becomes a very good use case. I can pretty much see entire sales people's first thing to do is basically have scouts. And that's why I was asking about once you get the scout, you want to put that data in something and then continue to track it. So that's why like even all my use cases, whenever I set up the next step is okay, I need to copy and paste this somewhere.

Abhishek Das (36:49.935)

Yeah, that's actually, yeah, it's like, it's a good thing and a bad thing because like, and I promise we are doing more on that front that's going to ship soon. But like we've heard this feedback from others as well that like, yeah, I'm looking at the results from Scouts. These are way better than other tools I've been using, but it doesn't quite have the look and feel and integrations that an enterprise or a B2B product should have. And so I have to like manually now copy paste stuff to like either Slack where it might get visibility in my team.

or copy paste it to like a spreadsheet or a CRM where I can use this for like the rest of my funnel. So yeah, like we're working on it. It's coming, but at least people seem to like the results.

Nataraj (37:33.903)

Yeah, I think you found some use case which is unique and you also found an abstraction level in which we interact with the web that is slightly different and unique that we haven't seen anywhere before. The closest thing I could think of like is a Google alert.

Abhishek Das (37:52.43)

Yeah, it is kind of like Google alert on steroids. that decision, that choice was very intentional because we, we, we, like, there's a lot of noise out there around agents, right? Like everyone's saying that, yeah, here's an agent that can do anything. And then you try it once and it doesn't quite do the right thing or do what you intended. That's frustrating. We wanted to make sure we are not like the hundred and forced company that's making the same mistake. And so we were very intentional that, okay, like how can we.

best package the power of these agents while still handling a use case that is relatively low-ish risk. So it is read-only. It's not going to buy book reserve things on your behalf. But you still get to see, because it's a general purpose, you still get to see the power of the tech. And the next versions of it will let you do some downstream actions, write actions, not just read actions.

Nataraj (38:50.821)

And if you think about it, 80 % of what we do in web is actually read. And most of the time we actually consume web and not actually contribute to it. And there's also this interesting thing that is happening with interacting with web, think is a couple of things happening. One is the web is shrinking. And I think we just accelerated the web shrinkage by a couple of, I don't know, percentage points in the last two, three years.

Abhishek Das (38:54.852)

Yeah.

Abhishek Das (39:18.861)

I'm curious to know what you mean by that because I think more pipe coded websites are getting created every single day than ever before in history.

Nataraj (39:26.499)

I think GitHub is increasing, GitHub size is increasing. But I think the incentive to publish wipe-coded apps is decreasing. mean, people can create their own apps that they want to use. So then you have to create more self-hosting avenues for folks to just run five to 10 users. Because everyone, like you have your own point of view on what not taking is. maybe there might be only maybe 100 people in the world who.

really are that specific, right? And like I created a poker tracking app for our friend group and it's only us using it, right? No one else uses it. So what really I mean is the incentive to write something to that, because there's no model, there's no models that are actually sort of incentivizing more people to contribute to it, whether it's blog posts.

You know, just even like YouTube videos in some sense have some good incentive mechanism. Like you have a creator contribution ecosystem, which is obviously, again, it follows power law, but it has some incentive, but just putting information on that. think fundamentally post-CHAT GPT is actually very low. I don't, I can't really put a number on it per se, but I feel like if it was a hundred, it is now 50 probably, at least half of it.

And creating a new blog is just now, it's just hobby. You cannot write something. And that's why you see people only writing blogs on Substacks or Behavis, because they offer certain kind of network they will promote you to, but they'll not have an individual blog. They will not, if you have niche thoughts, you cannot really expect distribution from Come From. SEO is completely changed.

Even if.

Abhishek Das (41:24.207)

Yeah, think there is a nugget of truth there. I do think the web is ripe for a change, like an infra or ecosystem-wide change to that effect. If most people are consuming content on the web through asking questions to an LLM, and the LLM is just going to pull information from every end website, the incentives aren't quite set up correctly for people to publish.

their own personal blogs. But instead, if an agent that accesses someone's blog could identify itself that, I'm acting on behalf of this product, on behalf of this user, and here's Tencent as part of this access that your blog can have, then the incentives get aligned quite well.

Nataraj (42:14.393)

Positive side also, the incentive to write trash on the web is also reduced. So you cannot just game SEO and get more views and views and views, right? So there's also that incentive is also reduced. So now you don't do SEO manipulation that you would do otherwise. So right now, like your answer, so getting you a good answer for a good question, high quality is actually worth doing. So if you're good at really something,

Abhishek Das (42:18.703)

Yeah, exactly.

Nataraj (42:39.973)

really any niche thing, then it becomes really valuable. So whatever you're really good at and not in the 80%, but now you have to be at the top 95, top 5 % to actually make it to that section. So it's sort of like the power law has queued a little bit more towards the top. And I think that's what changed in the web. That is, that's what I think.

Abhishek Das (43:03.747)

Makes sense? Yeah.

Nataraj (43:05.649)

I think we're almost at the end of our conversation. Can you talk a little bit more about, know, how are you thinking about finding product market fit? I mean, I like the product, but I don't know what the stats are. But do you think you guys have product market fit? How are you integrating? What are your general thoughts about like generally finding product market fit at this point?

Abhishek Das (43:32.984)

Yeah, we're still a pretty small company and product. We are seeing good, healthy metrics so far. Like our M1 retention, for example, is 80%. And just even qualitative feedback has been, by and large, quite positive. People come back, try the product, and come back and tell us that, if all the agentic products I've tried, this one

like gave me the most joy, was the most delightful, had the least hallucinations, et cetera, right? Yeah, so metric, like some of those early indicators are looking good. One of the things that is top of mind for us is to increase the surface area for the product. Like right now it just does information monitoring for you. It's broad in how it does that, but it's still that particular use case. And our not start is to be, for these agents to be able to do any task on your behalf on the web.

And so there's a long gap between those two. So just chugging along and making progress on capabilities on product experiences that make progress towards that North Star. And you'll see more from us in the coming weeks and months. But that's basically what's top of mind in search for getting to as many users as quickly as possible.

Nataraj (44:51.973)

How are you guys doing? How are you finding new users?

Abhishek Das (44:56.449)

A lot of it is through word of mouth. Yeah, we do a few marketing things every so often. When Scouts became generally available, we did a big marketing announcement and video, and that got a decent amount of new signups and so on. But largely, it's through word of mouth. People try it. They love it. They tell their friends about it or colleagues about it. They come to the product and try it out and so on.

Nataraj (45:23.813)

Awesome. I think one of the things that I'm seeing is when you left 2024 Meta, I Meta changed a lot since then. A lot of AI researchers were hired and then a lot of AI researchers came out of it. What do you think about what's happening with Meta at this point?

Abhishek Das (45:35.599)

Thanks

Abhishek Das (45:48.386)

Yeah, it's been a while since I left left mirror, especially in today's day and age like things move so quickly. So I don't know if I have a hot take that I think it's it's

It's definitely interesting from the outside to see what's happening at Meta. The willingness to throw this amount of money to hire the best people from the ecosystem is unprecedented to quite an extent, and can only happen if it's a founder-led company who would have this kind of a conviction of what they want to go after. So I'm just curious to see how that experiment plays out.

Like it's still early in TBD labs or like MSL or whatever they're calling it. It's still early in terms of like the products and models that they put out. But I'm just curious to see how this, how this experiment plays out. A lot of our team who, not just the founders, but like other, other members of the team who joined us, were like, they joined us before all of this, all of this happened. Like the whole shake up in the AI org at Meta.

Nataraj (46:58.319)

So I guess we might expect Zuckerberg to bring soup to your head office and I'll be back.

Abhishek Das (47:04.687)

Yeah, I'm still waiting for my 100 million offer. Just kidding.

Nataraj (47:10.629)

The other interesting thing I thought about when I was using Scouts was pricing. When I see similar products out there, they're not exactly quite similar product, but the price based on usage versus you guys chose to do a fixed subscription. Was that a conscious choice or?

Abhishek Das (47:26.563)

Mm-hmm.

Abhishek Das (47:37.294)

Yeah, so it just comes down to who you're building for. Like for example, for our API product, it is usage-based pricing, right? And that is what is standard in the ecosystem you spend, you pay based on how many tokens you use. But if you're targeting prosumers or consumers, usage-based pricing basically means that the incentive to use the product more is not there.

If every time you use the product, you're told that, OK, this is how much we're going to charge you, that doesn't quite set up the right incentive. So that was the primary reason why it didn't make sense to us. Ideally, what we would want to have is outcome-based pricing. Did it get the job done for you or not? And for some verticals, for example, if you're building a customer support agent,

it's very clear cut, like measuring that is very clear cut. How many tickets did you close or something? But in the domain that we are in, it can be more subjective. Like it's harder to measure, did it actually get the job done or not? Which is why the ideal version of like outcome-based pricing is just hard to roll out. But like just thinking through some of the next set of capabilities that we were thinking of in the next version of Scouts or Yutori, we, yeah, like.

we might switch what the pricing looks like just to be able to reconcile the perceived value and the cost of the product better. It is a weird word right now where even if, for example, we don't increase the number of use cases that Scouts does for you, but six months later, there's a better model. And so all your Scouts as a result get better and it costs us or

whoever we are getting the model from more to serve that product, there's a bit of a gap in that cost versus the perceived value. And a fixed monthly subscription doesn't quite reconcile that. So it's on our minds. We might change it in the future. Our users will, of course, get ample heads up whenever that change happens. But that's sort of the rationale and the thinking behind why we went with what we did.

Nataraj (49:55.896)

One last question I wanted to ask is because your background is in experimenting with robotics and near-phD, how do you see the whole robotics space shaping up?

Abhishek Das (50:09.121)

Again, like it's super interesting. Right now it feels like there are methods to generate synthetic data at an unprecedented scale for robotics that wasn't possible earlier. And so things are starting to finally work with, yeah, like which wasn't quite possible earlier. And that's why there's a ton of excitement around it. Part of our motivation around building Utori was that physical agents will happen eventually.

But it's probably unquestionable that digital agents should happen before physical agents become reality. Physical agents are just harder for a variety of other reasons. have other risks around them. But yeah, it is very exciting what's happening in physical agents. Someone recently was showing me a demo of using, I think, NanoBanana, or one of the image generation models, to translate a whole bunch of

images from the internet into first person images that you could, egocentric images that you could then use to train a brain for a robot. And it looked pretty good. And so if that's what I mean by if like there are scalable methods for generating data for robotics, that kind of changes the game and like time to do pre-eution for some of these ideas.

Nataraj (51:31.193)

I think we could probably like, Google probably can take all the images they have and like generate synthetic data at scale. Like that would look crazy. And it's also more like, I think those images will have more information about the physics of the real world than something if you just conjure up and might not really have. So I think it's still have to base it on some kind of.

reality and then sort of extrapolate it. But yeah, I think this was a great conversation. Abhishek, thanks for coming on the show and talking about your story. I'm super excited on what is next build. I'm with the consumer of the API for sure and looking forward. So if you have a beta program or an alpha program, feel free to include me.

Abhishek Das (52:22.511)

It's actually live on the website. If you go to utorii.com slash API, there's instructions there on how to sign up.

Nataraj (52:28.417)

I didn't realize that I was just using the other day. I'll definitely check it out. Thanks for coming.

Abhishek Das (52:34.072)

Awesome. Please keep Yeah, please keep sending us feedback. And yeah, thanks for having me.