This is the Biggest Hidden Risk of AI | Traversal, Anish Agarwal

1. 0:00 – 1:24

   Intro

   1. AAAnish Agarwal0:00

      The code is no longer being written by human, but by AI system. So much more code is being written now by Cursor or Windsurf or Copilot. AI is gonna write so much more code. No one really understands all of it. No one has full context. No team has full context about what's happening because it's such a complex system. When software breaks, it's gonna be really difficult to troubleshoot it, right? One of the biggest problems is downtime, is troubleshooting. Global IT cyber outage and- Global outage. Major IT outage. It's also affecting hospitals, law enforcement departments, banks, and major airlines. The cost of downtime to annually for all enterprises around $400 billion, so it's a huge problem. It could be several hours or days before the situation is fully resolved. Fundamentally, what we're doing is when we have large complex software systems, we help figure out-- and they break, we help figure out what happened. Having a team of on-call engineers twenty-four/seven looking at all of your data. So by the time an engineer comes on to a Slack channel, the root cause or candidate root cause is already given. So rather than them spending so much time trying to figure out what happened or, you know, calling more and more teams, 'cause typically what happens is you'll have one team look at the, the, the data. They're like, "Oh, it's not my fault." Then they'll call another team and another team and another team. That's how you go from five people to, like, eighty people on a channel, right? And so then rather than fifty people over an hour, it's like five to ten people for a few minutes just verifying the answer.

2. 1:24 – 2:32

   The problem I'm solving

   1. AAAnish Agarwal1:24

      My name is Anish. I'm the CEO and co-founder of Traversal. We came ourselves with forty-eight million dollars in seed and Series A funding, which were led by Sequoia and Kleiner Perkins. We just announced our Series A raise, and we came out of stealth. We're building an AI site reliability engineer. So what that means is when you have large complex software systems and they break, we troubleshoot to help you figure out why it broke and then help fix it automatically. Right, so it's like Perplexity. When, when you a-ask Perplexity a question, it kinda gives you evidence, right? It gives you citations of how it got to that answer. But our world, the, the citations aren't web links. The citations are links to your observability system. Now we can now publicly talk about, we've been working with them for the last six months now very closely, is DigitalOcean. So they're a large public cloud service provider. I think they're the third largest, actually, by number of people using them as their cloud provider. I think they have over six hundred thousand people, um, using them as their core infrastructure. And you can imagine when they break, you know, every person that is using them feels the pain, 'cause that's the main thing powering their, their system. We've been working with them for six months now, and we found that in that six-month period, we've dropped the, the time to resolution by over forty percent, like thirty-seven percent to be exact, which is incredible, right? Because as I said, every minute of downtime is, like, is worth thousands if not millions of dollars.

3. 2:32 – 4:23

   The Chat GPT Moment Changed Everything

   1. AAAnish Agarwal2:38

      I like sports. I like competing a lot, and I always naturally gravitated to math and science. I just liked how abstract and clean it was, and it felt quite universal in, in what you could do. I came to MIT just because I thought machine learning and AI was really important, and I wanted to understand it really deeply. This is like twenty sixteen, so about, like, eight, nine years ago. And once I got there, I think the-- one of the biggest moments in my, like, academic research career was NeurIPS twenty seventeen, and the keynote was given by the Google AlphaGo team. And I just found it incredible that a system can learn this creative thing by itself, because the complaint we always had before was that it's just copying people, but this was purely-- it was learning creativity by, by self-play. So I was like, "We should apply that everywhere, this kind of architecture." I was fortunate enough to get into Columbia's faculty. I like thinking about theoretical problems, mathematical abstractions, and I think that university is an amazing place to do that. The thing that changed-- And that was the time when everything with Chat GPT was, was happening, and so it just felt like something incredible has happened in the world, and it's like a once in a lifetime thing where the world has fundamentally changed. People don't even realize it. It just felt like this almost religious experience as to what was happening in the world. I really like uncertainty. I like creating something from zero to one. Similar between research and entrepreneurship is that the uncertainty. You have no idea what's happening most of the time, and you have to find ways of, of creating structure from nothing. I think obviously the difference is in here, the time spans are compressed, right? In research, you get five years, ten years to, to make an impact. Here you get one month, right? So the feedback cycle is very quick. But I think in this age of AI, when AI is changing so quickly, being in that quick feedback cycle is actually very important. We've kinda entered into the industrial age of artificial intelligence. You know, I also saw some of the smartest people around me. They were either at OpenAI or Anthropic or, you know, Meta, or they were creating companies, and that's what makes me really excited. And so I think starting a company was-- felt to me like a great

4. 4:23 – 7:27

   Begin with Your Edge

   1. AAAnish Agarwal4:23

      expression of that, I guess. I think the best AI companies are always gonna be at the edge of where the models are gonna be, right? That's how you differentiate yourself is you're always at the edge. If you're the edge, then sometimes it works, sometimes it doesn't work, and you need to know quickly when it's working and when it's not working and correct for that. When we started the company in, like, Jan of two thousand twenty-four, we started without an idea. But we had a clear taste of the type of problem we wanted it to take on. So we wanted to do something that was at the intersection of our research, which was in causal machine learning and reinforcement learning, uh, and how it intersected with AI agents. Causal machine learning is the study of cause and effect, and what you want to understand is how do you get these AI systems to pick up cause and effect relationships from data. A/B test is an example of, of learning cause and effect relationships. A clinical trial is another example. So these are, like, basic ways of, of running experiment. And so that's what the study of, of causal machine learning is. And we're trying to see how did that intersect with AI agents, uh, which we thought was, like, super cool and something we'd followed for, like, now almost two and a half years. We went through a few different ideas. The fourth person who joined us, our fourth co-founder, Amir, and so he pitched us the problem of dealing with incidents. And as we looked into it, it kind of felt like a perfect problem, finding this needle in a haystack with many fake needles everywhere. So it fit with our research in causal machine learning and reinforcement learning really well. It fit with LLMs really well because the haystack is composed of, like, logs and metrics and traces and code and configuration files and so on and so forth. It fits with AI agents really well because you have to automate this complex workflow where you're, you know, querying all these different pieces of software, you're reasoning over them, and then you're writing more queries, and it's like this, like, sequential adaptive flow. And so it's a big market because everyone cares about software not going down. And I think it's only gonna get bigger, right? Because so much more code is being written now by companies like Cursor or Windsurf or Copilot or what have you. Just like Cambrian explosion of code being written. No one understands it in some ways. And so when software breaks, it's, it's gonna be really difficult to troubleshoot it, right? And so I think that's what gave us confidence that this is the problem we should be taking on. And then honestly, the first VC I met in my life was Sequoia, and I think they have been looking for a, a, a team to solve this problem. They had reached the same thesis. They felt this is a problem where like the AI risk and technical risk is really high and the market risk is low because if you can solve it, there's a big market. And so people like us who don't come from this world but are good on the AI side are the right people to solve it. And so obviously that validation also gave us confidence that this is the right problem. A lot of what these AI agent companies are doing is trying to replicate what humans have done, but there's so much more that can be done. And so thinking from first principles what AI systems are good at and exploiting that versus just trying to replicate what a human has done, I think is also gonna be very important to reinvent and actually get to the next level of innovation. Creating a, a MVP is so easy with all the tools out there. You can really iterate quickly with putting a product in people's hands. We built our first MVP probably last year in June or July, like about three months in. And with small companies it worked great. Because the scale of the data was small, we could kinda look at their historical incidents, see what the playbook was, and then put that into an AI agent. And so like our accuracy was like 90%, something amazing, right? So

5. 7:27 – 9:42

   The First Principle Saved Us From 0% Accuracy

   1. AAAnish Agarwal7:27

      we felt really confident that this is gonna work. Just because it works in the one time doesn't mean it's always gonna work because the world is constantly changing. And then we hit some of the larger enterprises including DigitalOcean. Our accuracy went to 0%, which is very difficult [laughs] to see. It was a tough week. Creating an MVP is easy, but creating a production system that works in complex environments is really hard. And so I think once you're not confused an MVP with a production AI system, those are like two very, very, very different things. But then we re-architected a lot of things. We said, "How do we make sure that we're no longer trying to use our creativity and seeing, you know, get that into an agent, but really use what these AI systems are good at, which is using computation, right? Using inference." That's really what unlocked us, and suddenly our accuracy went back up to 90%. How do you make sure that your system gets better with the reasoning models? 'Cause there are a lot of people we saw in competing companies and so on and so forth, where as the reasoning model came out, they didn't get any better. So how do you make sure that you're exploiting what the reasoning models are good at? And the way I put it is that the reasoning models are very good at like detective stories. You have a mystery novel and you're trying to figure out who, who did the crime. And there's all these different pieces of evidence that you're seeing and you're trying to figure out who was the person who did it. Connecting all those dots and figuring out the thing, you know, the person who did it, that kind of detective story type workflow is what I think these reasoning models are, are very good at, where you have a clear answer at the end and you have lots of moving pieces that you have to like connect the dots between to get to the clear answer. That felt kind of perfect for us, right? Because for us, we have all these different symptoms that are happening at the same time, and you find ways to connect the dots to find that specific right answer, like who did it. Making sure we were exploiting them to the maximum was, was crucial I think. And so I think that's the way I'd put it. AI is gonna write so much more code and no one really understands all of it, right? If you wrote all of it, you have in your head just how it all fits. And as you get to bigger and bigger systems already, right, you work with some of the largest Fortune 100 companies, no one has full context, no team has full context about what's happening because it's such a complex system. And that's just happening not just at the large companies, but also at the small companies because the s- the code is no longer being written by a human, but by-- or an engineer, but by AI system, right? So the lack of context means that when it's, when an incident happens, it's just so much harder to debug it because you just don't have all of the context you need. And actually it's already happening, like most people now are not developing code, they're starting

6. 9:42 – 12:45

   How to Survive the AI Coding Era - Do What You Love with People You Love

   1. AAAnish Agarwal9:42

      to like validate QA code or troubleshoot code, and that doesn't scale. That's the big problem. And I think the second big problem I think is that, you know, with all the, all the developments happening with AI software engineering, our belief is that as engineers we get to do the really creative fun work, architecting system design. Over time, all engineers will be doing will be troubleshooting, which will be sad in my opinion. Like they should be doing the most-- We as a, as engineers should be doing the most creative work, right? And to, to make that a reality, you, you need to have systems not just developing a software and building it, but also maintaining it. And so I think all the software maintenance needs to be reinvented. You have to just kind of persevere. If something goes wrong, that's okay. I think it, in some ways it's a good thing because if it was easy, then everyone could do it, right? I think this is one of those problems where the problem statement is very easy to state, but to actually solve it is really hard. And I think that's why there's like a lot of companies trying it, but very few actually succeeding. And I think having the ability to like stay resilient and have grit when something doesn't work and just stick with the problem is I think a big part of what differentiates us, uh, as a company. Think about what's gonna be important 10 years from now, regardless of whatever happened in the world. You have no idea what's happening most of the time, and you have to find ways of, of creating structure from nothing. And so like one way I say it is that, you know, in typically hard jobs, whether it's in finance or it's in technology as an engineer, like you have a point A, you figure out point B, and it's very hard to get from point A to point B. In the world of research and also in the world of entrepreneurship, you don't really know where point A is. You don't know where you are. You don't know where point B is. You don't know where you wanna go. And if you did know, it's still very hard. And so you're constantly like in this game of trying to just decide where you are and where you're trying to go, and that's exactly the same thing in research as well. What's interesting in this job is that the, the stresses are, are, can be very high, lows can be very low. So I think getting used to like just very high highs and low lows is important. But I think the one thing I've learned is that the time it takes me to, to recover is very fast. Even if I'm super stressed and super tired, within one or two days if I just take it off, I'm back to full force. And so I think that's been like a good learning, is that if you're really enjoying what you're doing, and even though there's these moments of massive stress, levels of stress that you'll never face otherwise, if you really love a problem and you're, you're attracted by it, that's where it attracts other people, right? We all love the problem we're all facing in our lives, and so we really feel like a great deep desire to solve it. And probably the most important thing out of anything is surrounding yourself with people that you care, that you like, that you wanna be like. And if you do that, life will be okay. I think about my research life, I found the right PhD advisor or advisors that really molded me and guided me in the right way. If I think about this role, I found the right investors that guided me at the start, the right customers that, you know, that helped define the product. You live and die by the people you, you surround yourselves with, and the people who will kinda guide you through these different parts. And so I think building a taste for the right people to mentor you and, and be a partner with you is the most important thing. The rest of it I think will figure itself out if you can find the right people around you. But I wouldn't just start a company for the sake of it. I think you should, you should feel like something deep in you, 'cause it's not easy. [laughs] And so you need that kinda deep belief or, you know, or motivation to sustain you over time. [outro music]

Episode duration: 12:57