Transcriptly

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

EO12:57Added Aug 31, 2026

"Human Coding is Dead," said Anish Agarwal, cofounder of Traversal, in our interview. So much of today's software is already written by AI. Tools like Cursor...

Watch on YouTube →
Contributed by 刘嘉琪

Transcript

Transcript format

  1. Intro

  2. 00:00The code is no longer being written by
  3. 00:01human but by AI system. So much more
  4. 00:04code is being written now by cursor or
  5. 00:06Windsor for copilot. AI is going to
  6. 00:08write so much more code. No one really
  7. 00:10understands all of it. No one has full
  8. 00:12context. No team has full context about
  9. 00:13what's happening because it's such a
  10. 00:15complex system. When software breaks,
  11. 00:17it's going to be really difficult to
  12. 00:18troubleshoot it. Right? One of the
  13. 00:19biggest problems is downtime is
  14. 00:21troubleshooting. Global IT cyber outage.
  15. 00:24Global outage, major IT outage.
  16. 00:26It's also affecting hospitals, law
  17. 00:28enforcement departments, banks, and
  18. 00:30major airlines.
  19. 00:31The cost of downtime to annually for all
  20. 00:34enterprises is around $400 billion. So,
  21. 00:35it's a huge problem.
  22. 00:36It could be several hours or days before
  23. 00:39the situation is fully resolved.
  24. 00:41Fundamentally, what we're doing is when
  25. 00:43we have large complex software systems,
  26. 00:45we help figure out and they break, we
  27. 00:47help figure out what happened. having a
  28. 00:48team of on call engineers 24/7 looking
  29. 00:51at all of your data. So by the time an
  30. 00:53engineer comes onto a Slack channel, the
  31. 00:55root cause or current root cause is
  32. 00:57already given. So rather than them
  33. 00:59spending so much time trying to figure
  34. 01:01out what happened or you know calling
  35. 01:02more and more teams because typically
  36. 01:04what happens is you'll have one team
  37. 01:05look at the the the data. They're like
  38. 01:07oh it's not my fault. Then they'll call
  39. 01:08another team and another team and
  40. 01:10another team. That's how you go from
  41. 01:11five people to like 80 people on a
  42. 01:13channel, right? And so then rather than
  43. 01:1450 people over an hour, it's like 5 to
  44. 01:1710 people for a few minutes just
  45. 01:18verifying the answer.

  46. The problem I'm solving

  47. 01:24My name is Anish. I'm the CEO and
  48. 01:26co-founder of Traversal. We came
  49. 01:27ourselves with $48 million in the series
  50. 01:29A funding which were led by Sequoia and
  51. 01:31Kleiner Perkins. We just announced our
  52. 01:33series A raise and we came out of
  53. 01:35stealth. We're building an AI site
  54. 01:36reliability engineer. So what that means
  55. 01:38is when you have large complex software
  56. 01:40systems and they break, we troubleshoot
  57. 01:41to help you figure out why it broke and
  58. 01:43then help fix it automatically. But it's
  59. 01:44sort of like perplexity when when you
  60. 01:46ask perplexity a question, it kind of
  61. 01:47gives you evidence, right? It gives you
  62. 01:48citations of how it got to that answer.
  63. 01:50But our world that the citations aren't
  64. 01:52web links, the citations are links to
  65. 01:54your observability system that we can
  66. 01:56now publicly talk about. We've been
  67. 01:57working with them for the last 6 months
  68. 01:58now very closely is Digital Ocean. So
  69. 02:00they're a large public cloud service
  70. 02:03provider. I think they're the third
  71. 02:04largest actually by number of people
  72. 02:05using them as their cloud provider. I
  73. 02:07think they have over 600,000 people um
  74. 02:09using them as their core infrastructure.
  75. 02:11And you can imagine when they break, you
  76. 02:12know, every person that is using them
  77. 02:14feels the pain because that's the main
  78. 02:16thing powering their their system. We've
  79. 02:18been working with them for 6 months now
  80. 02:20and we found that in that 6 month period
  81. 02:21we've dropped the the time to resolution
  82. 02:24by over 40% like 37% to be exact, which
  83. 02:26is incredible, right? Because as I said,
  84. 02:28every minute of downtime is like is
  85. 02:30worth thousands of millions of dollars.

  86. The Chat GPT Moment Changed Everything

  87. 02:37I like sports. I like competing a lot
  88. 02:39and I always naturally gravitated to
  89. 02:41math and science. I just liked how
  90. 02:43abstract and clean it was and it felt
  91. 02:45quite universal in and what you could
  92. 02:47do. I came to MIT just because I thought
  93. 02:49machine learning and AI was really
  94. 02:50important and I wanted to understand it
  95. 02:51really deeply. This is like 2016. about
  96. 02:53like 8 n years ago and once I got there
  97. 02:55I think the one of the biggest moments
  98. 02:57in my like academic research career was
  99. 02:59Europe's 2017 and the keynote was given
  100. 03:02by the Google AlphaGo team and I just
  101. 03:04found that incredible that a system can
  102. 03:06learn this creative thing by itself
  103. 03:08because the complaint you always had
  104. 03:09before was that it's just copying people
  105. 03:11but this was purely it was learning
  106. 03:13creativity by by selfplay so I was like
  107. 03:15we should apply that everywhere this
  108. 03:16kind of architecture I was fortunate
  109. 03:18enough to get into Colombia's faculty
  110. 03:19and I like thinking about theoretical
  111. 03:21problems mathematical abstractions And I
  112. 03:23think university is an amazing place to
  113. 03:25do that. The thing that changed and that
  114. 03:27was the time when everything with Chad
  115. 03:28GPT was was happening. And so it just
  116. 03:30felt like something incredible has
  117. 03:32happened in the world. And it's like a
  118. 03:33once in a-lifetime thing where the world
  119. 03:35is fundamentally changed. People don't
  120. 03:36even realize it. It just felt like this
  121. 03:38almost religious experience as to what
  122. 03:39was happening in the world. I really
  123. 03:40like uncertainty. I like creating
  124. 03:42something from 0 to one. similar between
  125. 03:44research and entrepreneurship is that
  126. 03:46the uncertainty you have no idea what's
  127. 03:48happening most of the time and you have
  128. 03:50to find ways of of creating structure
  129. 03:52from nothing I think obviously the
  130. 03:53difference is in here the time spans are
  131. 03:57compressed right in research you get 5
  132. 03:59years 10 years to to make an impact here
  133. 04:01you get 1 month right so the feedback
  134. 04:03cycle is very quick but I think in this
  135. 04:05age of AI when AI is shooting so quickly
  136. 04:07being in that quick feedback cycle is
  137. 04:08actually very important we've kind of
  138. 04:10entered into the industrial age of
  139. 04:12artificial intelligence you And I also
  140. 04:13saw some of the smartest people around
  141. 04:14me. They were either at OpenAI or
  142. 04:16Enthropic or you know Meta or they were
  143. 04:18creating companies and that's what makes
  144. 04:20me really excited. And so I think
  145. 04:21starting a company was felt to me like a

  146. Begin with Your Edge

  147. 04:23great expression of that. I guess
  148. 04:30I think the best AI companies are always
  149. 04:32going to be at the edge of where the
  150. 04:33models are going to be, right? That's
  151. 04:34how you differentiate yourself is you're
  152. 04:35always at the edge. If you're the edge
  153. 04:37and sometimes it works, sometimes it
  154. 04:38doesn't work. And you need to know
  155. 04:40quickly when it's working and when it's
  156. 04:41not working and correct for that. When
  157. 04:43we started the company in like January
  158. 04:44of 2024, we started without an idea. But
  159. 04:46we had a clear taste of the type of
  160. 04:48problem we wanted to take on. So we
  161. 04:49wanted to do something that was at the
  162. 04:51intersection of our research which was
  163. 04:52in causal machine learning and
  164. 04:54reinforcement learning and how it
  165. 04:55intersected with AI agents. Causal
  166. 04:57machine learning is a study of cause and
  167. 04:59effect. And what you want to understand
  168. 05:00is how do you get these AI systems to
  169. 05:02pick up cause and effect relationships
  170. 05:04from data. AB test is an example of of
  171. 05:06learning cause and effect relationships.
  172. 05:07like clinical trial is another example.
  173. 05:09So these are like basic ways of of
  174. 05:10running experiment. And so that's what
  175. 05:12the study of of causal machine learning
  176. 05:13is and we're trying to see how did that
  177. 05:15intersect with AI agents uh which we
  178. 05:17thought was like super cool and
  179. 05:18something we followed for like now
  180. 05:19almost 2 and a half years. We went
  181. 05:21through a few different ideas. The
  182. 05:22fourth person who joined us our fourth
  183. 05:24co-founder Ahmed and so he pitched us
  184. 05:26the problem of dealing with incidents.
  185. 05:28And as we looked into it, it kind of
  186. 05:29felt like a perfect problem finding this
  187. 05:30needle in a hay stack with many fake
  188. 05:32needles everywhere. So it fit with our
  189. 05:34research in causal machine learning and
  190. 05:35reinforcement learning really well. It
  191. 05:37fit with LLMs really well because the
  192. 05:39haststack is composed of like logs and
  193. 05:40metrics and traces and code and
  194. 05:42configuration files and so on and so
  195. 05:43forth. It fits with AI agents really
  196. 05:45well because you have to automate this
  197. 05:46complex workflow where you're, you know,
  198. 05:48querying all these different pieces of
  199. 05:50software. You're reasoning over them and
  200. 05:51then you're writing more queries and
  201. 05:52it's like this like sequential adaptive
  202. 05:54flow. And so it's a big market because
  203. 05:56everyone cares about software not going
  204. 05:58down. And I think it's only going to get
  205. 05:59bigger, right? because so much more code
  206. 06:02is being written now by companies like
  207. 06:04cursor or windsurf or copilot or what
  208. 06:06have you just like cambrian explosion of
  209. 06:09code being written no one understands it
  210. 06:11in some ways and so when software breaks
  211. 06:14it's it's going to be really difficult
  212. 06:15to troubleshoot it right and so I think
  213. 06:16that's what gave us confidence that this
  214. 06:18is the problem we should be taking on
  215. 06:20and then honestly the first VC I met in
  216. 06:21my life was Sequoa and I think they've
  217. 06:24been looking for a team to solve this
  218. 06:26problem they reached the same thesis
  219. 06:27they felt this is the problem where like
  220. 06:28the AI risk and technical risk is high
  221. 06:30and the market risk is low because if
  222. 06:31you can solve it, there's a big market
  223. 06:33and so people like us who don't come
  224. 06:34from this world but are good on the AI
  225. 06:36side are the right people to solve it.
  226. 06:37And so obviously that validation also
  227. 06:39give us confidence that this is the
  228. 06:40right problem.
  229. 06:46A lot of what these AI agent companies
  230. 06:48are doing is trying to replicate what
  231. 06:50humans have done, but there's so much
  232. 06:51more that can be done. And so thinking
  233. 06:53from first principles what AI systems
  234. 06:55are good at and exploiting that versus
  235. 06:57just trying to replicate what a human
  236. 06:59has done I think is also going to be
  237. 07:00very important to reinvent and actually
  238. 07:02get to the next level of innovation.
  239. 07:04Creating a MVP is so easy with all the
  240. 07:07tools out there. You can really iterate
  241. 07:09quickly with putting a product in
  242. 07:10people's hands. We built our first MVP
  243. 07:12probably last year in June or July like
  244. 07:15about 3 months in. And with small
  245. 07:16companies it worked great because the
  246. 07:18scale of the data was small. we could
  247. 07:19kind of look at their historical
  248. 07:20incidents, see what the playbook was and
  249. 07:22then put that into an AI agent. And so
  250. 07:24like our accuracy was like 90%.
  251. 07:26Something amazing, right? So we felt

  252. The First Principle Saved Us From 0% Accuracy

  253. 07:27really confident that this is going to
  254. 07:28work. Just because it works in the one
  255. 07:30time doesn't mean it's always going to
  256. 07:31work because the world is constantly
  257. 07:32changing. And then we hit some of the
  258. 07:34larger enterprises including Digital
  259. 07:35Ocean, our accuracy went to 0%. Which is
  260. 07:38very difficult to see. It was a tough
  261. 07:39week. Creating an MVP is easy, but
  262. 07:41creating a production system that works
  263. 07:43in complex environments is really hard.
  264. 07:45And so I think one should not confuse an
  265. 07:48MDP with a production AI system. Those
  266. 07:50are like two very very very different
  267. 07:52things. But then we rearchitected a lot
  268. 07:54of things. We said how do we make sure
  269. 07:55that we're no longer trying to use our
  270. 07:58creativity and seeing you know get that
  271. 07:59into an agent but really use what these
  272. 08:01AI systems are good at which is using
  273. 08:03computation right using inference.
  274. 08:05That's really what unlocked us and
  275. 08:06suddenly our accuracy went back to up to
  276. 08:0890%. How do you make sure that your
  277. 08:10system gets better with the reasoning
  278. 08:13models? because there are a lot of
  279. 08:14people we saw in competing companies and
  280. 08:16so on and so forth where as the reason
  281. 08:18models came out they didn't get any
  282. 08:19better. So how do you make sure that
  283. 08:20you're exploiting what the reasoning
  284. 08:22models are good at? And the way I put it
  285. 08:23is that the reasoning models are very
  286. 08:24good at like detective stories. You have
  287. 08:26a mystery novel and you're trying to
  288. 08:27figure out who who did the crime.
  289. 08:29There's all these different pieces of
  290. 08:30evidence that you're seeing and you're
  291. 08:31trying to figure out who is the person
  292. 08:32who did it. Connecting all those dots
  293. 08:34and figuring out the thing, you know,
  294. 08:35the person who did it. that kind of
  295. 08:36detective story type workflow which I
  296. 08:39think these reasoning models are very
  297. 08:40good at where you have a clear answer at
  298. 08:42the end and you have lots of moving
  299. 08:43pieces that you have to like connect the
  300. 08:44dots between to get to the clear answer
  301. 08:46that felt kind of perfect for us right
  302. 08:48because for us you have all these
  303. 08:49different symptoms that happen at the
  304. 08:50same time you find ways to connect the
  305. 08:51dots to find that specific right answer
  306. 08:53like who did it making sure we were
  307. 08:55exploiting them to the maximum was was
  308. 08:57crucial I think and so I think that's
  309. 08:58the way I would put it
  310. 09:04is going to write so much more code and
  311. 09:07no one really understands all of it,
  312. 09:08right? If you wrote all of it, you have
  313. 09:10in your head just how it all fits. And
  314. 09:12as you get to bigger and bigger systems
  315. 09:13already, right, you work with some of
  316. 09:15the largest fortune 100 companies, no
  317. 09:17one is full context. No team is full
  318. 09:18context about what's happening because
  319. 09:20it's such a complex system. And that's
  320. 09:22just happening not just at the large
  321. 09:23companies, but also at the small
  322. 09:24companies because this the code is no
  323. 09:26longer being written by human but by or
  324. 09:29engineer but by AI system, right? So the
  325. 09:31lack of context means that when it's
  326. 09:33when an incident happens, it's just so
  327. 09:34much harder to debug it because you just
  328. 09:36don't have all of the context you need.
  329. 09:38And actually it's already happening like
  330. 09:40most people now are not developing code.
  331. 09:41They're starting to like validate QA

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

  333. 09:43code or troubleshoot code and that
  334. 09:45doesn't scale. So that's the big
  335. 09:46problem. And I think the second big
  336. 09:47problem I think is that you know with
  337. 09:49all all the developments happening with
  338. 09:50AI software engineering, our belief is
  339. 09:52that as engineers we get to do the
  340. 09:54really creative fun work architecting
  341. 09:56system design. Over time, all engineers
  342. 09:59will be doing will be troubleshooting,
  343. 10:00which would be sad in my opinion. Like
  344. 10:02they should be doing the most we as a as
  345. 10:04engineers should be doing the most
  346. 10:05creative work, right? And to to make
  347. 10:06that a reality, you you need to have
  348. 10:08systems, not just developing your
  349. 10:10software and building it, but also
  350. 10:11maintaining it. And so I think all of
  351. 10:13software maintenance needs to be
  352. 10:14reinvented. You have to just kind of
  353. 10:16persevere. If something goes wrong,
  354. 10:17that's okay. I think it in some ways
  355. 10:19it's a good thing because if it was
  356. 10:20easy, then everyone could do it, right?
  357. 10:22I think this is one of those problems
  358. 10:23where the problem statement is very easy
  359. 10:25to state, but to actually solve it is
  360. 10:26really hard. And I think that's where
  361. 10:28there's like a lot of companies trying
  362. 10:29it, but very few actually succeeding.
  363. 10:30And I think having the ability to like
  364. 10:32stay resilient and have grit when
  365. 10:34something doesn't work and just stick
  366. 10:35with the problem is, I think, a big part
  367. 10:37of what differentiates us uh as a
  368. 10:38company. Think about what's going to be
  369. 10:40important 10 years from now regardless
  370. 10:41of whatever happened in the world. You
  371. 10:43have no idea what's happening most of
  372. 10:45the time. And you have to find ways of
  373. 10:46of creating structure from nothing. And
  374. 10:48so like one way I say it is that you
  375. 10:50know in typically hard jobs whether it's
  376. 10:52in finance or it's in technology as an
  377. 10:55engineer like you have a point A if you
  378. 10:57get a point B and it's very hard to get
  379. 10:58from point A to point B in the world of
  380. 11:00research and also in the world of
  381. 11:01entrepreneurship you don't really know
  382. 11:02where point A is you don't know where
  383. 11:03you are you don't know where point B is
  384. 11:05you don't know where you want to go and
  385. 11:06if you did know it's still very hard and
  386. 11:07so you're constantly like in this game
  387. 11:09of trying to just decide where you are
  388. 11:11and where you're trying to go and that's
  389. 11:12exactly the same thing in research as
  390. 11:13well. What's interesting in this job is
  391. 11:16that the the stresses are are can be
  392. 11:18very high, lows can be very low. So, I
  393. 11:20think getting used to like just very
  394. 11:21high highs and low lows is important.
  395. 11:24But I think that one thing I've learned
  396. 11:25is that time it takes me to to recover
  397. 11:27is very fast. Even if I'm super
  398. 11:29stressed, I'm super tired, within one or
  399. 11:30two days, if I just take it off, I'm
  400. 11:32back to full force. And so I think
  401. 11:33that's been like a good learning is that
  402. 11:34if you're really enjoying what you're
  403. 11:36doing and even though there's these
  404. 11:37moments of massive stress, levels of
  405. 11:39stress that you'll never face otherwise,
  406. 11:40if you really love a problem and you're
  407. 11:42you're attracted by it, that's where
  408. 11:44attracts other people, right? We all
  409. 11:46love the problem. We're all faced in our
  410. 11:48lives. And so we really feel like a
  411. 11:50great deep desire to solve it. And
  412. 11:52probably the most important thing out of
  413. 11:53anything is surround yourself with
  414. 11:55people that you care that you like that
  415. 11:56you want to be like. And if you do that,
  416. 11:58life will be okay. I think about my
  417. 11:59research life. I found the right PhD
  418. 12:01adviser advisers that really molded me
  419. 12:04and guided me the right way. If I think
  420. 12:06about this world, I found the right
  421. 12:07investors that guided me at the start,
  422. 12:09the right customers that you know that
  423. 12:11helped define the product. You live and
  424. 12:13die by the people you you surround
  425. 12:15yourselves with and the people will kind
  426. 12:16of guide you through these different
  427. 12:17parts. And so I think building a taste
  428. 12:19for the right people to mentor you and
  429. 12:21and be a partner with you is the most
  430. 12:23important thing. The rest of it, I
  431. 12:24think, will figure itself out if you can
  432. 12:25find the right people around you. But I
  433. 12:27wouldn't just start a company for the
  434. 12:28sake of it. I think you should you
  435. 12:29should feel like some thing deep in you
  436. 12:32cuz it's not easy. And so you need that
  437. 12:34kind of deep belief or you know or
  438. 12:37motivation to sustain you over time.