Why AI Agents Fail in the Real World | Temporal on Lightwork

0:01Congratulations. As of this week today, Temporal has raised a$550 million Series E at a$12.55 billion valuation. Actually, just seven months after around that value, this company at$5 billion. How are you both feeling? Super excited, super proud of the team. And like, it's been an awesome moment for us to get here. Likewise, it's been kind of an incredible journey to work with Summer and the team since we last invested at this round and then, you know, having the chance to double down this go around. Before we get into the round and what this round entails, I actually want to back up a little bit for those who might not be as familiar and not know what the company does.

0:45It might be helpful to just high level, help me understand what does Temporal do and help me understand, you know, in my day to day, whether it's Chat2BT or paying a bill, how it shows up. Yeah, absolutely. I can give it a shot. So Temporal is an open source developer platform which ensures doable execution of your code. What that means is imagine your favorite food delivery, like a website or app, right? When you place an order as a customer, it's eventually picked up by a restaurant, goes to a kitchen where they work on your order, prepare the food. Then a driver shows up, picks it up and delivers it to the customer.

1:29So imagine like while this entire process is running, it goes through a lot of transitions from one state to another. And on the surface, it seems super straightforward. But what I have seen as a developer again and again, developers are building these insane systems to guarantee these things can never fail in the middle. They are using low-level primitive like queues and retries and databases and timers to kind of build resiliency and to guarantee that each and every order shows up at customer's doorsteps and the customer being charged. Imagine like as a customer, like you placed an order and then while the restaurant is working on it and something bad happens, machine restarts or a process crashes, you lose all your state.

2:20And now you have a restaurant which is super upset because they prepared a food and no one came to pick it up. And what's even worse is you have a really hungry customer. And this is where the engineering has been building those backend systems to guarantee that reliability rather than focus on your applications. And this is exactly the toil temporal removes, where we give you that reliability substrate, which allows the developer to build and deliver value in a reliable fashion to their customer. So it sounds like reliability, massive pain point, clearly, as you've laid out. So Anushka, have you been spending time on the market?

2:58Obviously, it sounds like we identified there was a problem here, clearly, in the way that this team also identified. How did you think about this specific company and their solution to it all? So one of our core beliefs as a firm and what we're seeing play out is that AI is going to handle a lot of work. And it's not just about making your software a little smarter, but it's actually about producing agents that do the work. We're seeing all these incredible AI experiences right now, like booking your flight, shipping code, making a payment. And for agents to do all of those things, you know, two things need to be true.

3:32They need to be reliable and they need to be long running. So reliable means that they can touch real business systems and the outside world, call a payments API, update a record, you know, finish a job correctly. And these agents need to work for hours and days at a time. And so if you take basically any AI experience you love right now, there's a decent chance that companies are not running temporal under the hood. And we've been so lucky to have been on the journey with Summer and team. Tell me a little bit more about where you think these models fail. And it sounds like it's not because they can't handle the real world execution.

4:07And tell me a little bit more about how you realize like when a demo isn't going to actually survive production. Yeah, I think it's less about a model failing. it's about like these models are getting smarter and smarter and as these models get smarter to Anushka's point earlier that's like now people are using it to take real action in the in the real world so and we hear that again and again from our users that the longer an agent runs the more tools it invoke the more actions it takes the more value it creates and basically once you are living in this world is basically your chances of failures compound as the longer this agent runs it the more systems it's touching and each one of those systems can have its own failure modes and it's compounding the chances of the failures essentially so and this is exactly where temporal becomes like a substrate where it abstracts out it becomes execution authority because model is no longer the constraint.

5:16We know these models are getting smarter and smarter with every new release that comes out. But then the bottleneck that it becomes, this is exactly where these demos fail is now when these agents start taking actions in the real world and their inability to handle those failures because real world is complex, like either it's waiting for a human, which might take two days to respond back. Or there is like you do a cancellation to a downstream system. But after the cancellation, it's not responded back to the user. And you get basically, you do the refund and the response doesn't come back. And then like if a failure happens, what do you do?

5:57Do you retry the whole cancellation again and do another refund? That would be really bad for the actual business because doing refunds twice. So what we see again and again that there are like infinite number of those failure modes. And this is exactly where which becomes a bottleneck for those agents that they hit the real world. That's exactly it. Like, you know, if you think about the last generation of software, you know, failure wasn't great, but it was OK in the sense that you were sitting in front of a computer and I could keep hitting approve. And if something doesn't work, I just hit retry.

6:32And with this new generation of software, I'm telling my agent to go book me a flight or update my bank details. So, you know, if the agent comes back in a few hours with no results, I might be unhappy. If the agent gets something like a payment wrong, I'm completely furious. And so this is why, you know, people are reaching for temporal at the outset. Yeah. And it's so interesting because there's just so much behind the scenes plumbing that you don't realize every single step becomes like a failure mode. And I'm just thinking through even the checkout flow, how many times an agent could have potentially failed for me.

7:03For a developer, what does this look like? How does it show up differently once you're using something like a temporal? so i think that this is where it the first time a developer sees something like devil a temporal like it completely blows up their mind essentially because a programming model imagine a subscription workflow a subscription workflow is basically every 30 days when a month ends you want to check if the user is still subscribed to a service and if they are like you want to charge that user for the subscription it's as simple as that but building those systems requires orchestrating like you need to have a timer service to remind you 30 days have passed.

7:43You need to call to a database which could still fail to see if the user still is a subscriber for the service or not. So it requires a bunch of those things. With temporal, you just write four lines of code in a loop. You basically, you say sleep 30 days. And when the timer fires, after 30 days, you check subscription and charge the user and go back to the loop. Just four lines of code. and you put those four lines of code in front of a developer, their mind is blown, like, because it's very counterintuitive for them, is like, oh, can I really write the four lines of code as production? And you can keep on running that group forever.

8:21And then we handle all sorts of failures underneath the cover. And that is so freeing to a developer where they now, once they get this model, oh, I just write a function and this function will run to its completion with all sorts of failures and nothing can interrupt that. then it completely changes their model of designing the next generation of systems. Anushka, I know you talk to a lot of AI teams. And so when you had spoken to some of them that had used Temporal, did you see that sort of like aha moment or pattern? Like, tell me more about how you realized that this was the team that was going to unlock that moment.

8:57Yeah, I mean, it's insane. The thing that was also, you know, remarkable to hear is you talk to some of these teams at the labs and whatnot, and their response would be, hey, Temporal is the only game in town. And so one of the things that I think is so cool about Temporal is you have sophisticated AI companies using the technology and then you have these two-person startups realizing it and using it too. And I think one of our big takeaways from all of the work is that there are very, very few infrastructure companies truly linked to the token flow. and temporal is one of them, where as your agent does more work, you know, they consume more temporal actions.

9:42Tell us more about some of the more interesting temporal use cases, since it sounds like there's just so many diversified use cases here. Yeah, so there are like thousands and thousands of use cases on top of temporal. One of the things that I, by the way, love the most is a bridge, which is healthcare. I can personally relate to it is it's so like, like freeing is whenever I take my mother to her appointment in the past, what I have seen is the, the doctor is paying more attention to writing the notes than actually really talking to the patient. And sometimes really just all the patient needs is someone who pays attention to them and then listens to them and then have a conversation.

10:30And I think this is why I love the problem that Abridge or like healthcare companies are solving. At the end of the day, it's like coordination of high volume, AI transcription, summarization, clinical notes. And it's been a game changer. And it's just like has really lifted the experience for a patient when they are in those conversations with a doctor, where now the doctor can pay full attention to them rather than taking those notes so they can hand it off to someone for rest of the processing, essentially. And I love that use case because I personally relate to it. And the role that temporal plays over there is that each and every one of those conversations, this is your medical, this is your health.

11:16Something like, oh, it usually works or sometimes works is not a good enough answer. So Temporal is the reliability backbone for the entire transcription pipeline, summarization, clinical note-taking, and then post-processing. With all of those things that happens in the background, Temporal provides the guarantees across that entire pipeline. I actually think one of the tidbits I had heard was that this was sort of the fourth or fifth time integrating on a system like this across your time at Amazon, Azure, Uber, Cadence. Tell me a little bit more about why this was sort of the right moment to go build this again and what Temporal finally got right this time.

11:59Yeah, so actually the interesting thing is we started the company back in 2019. And to your point, the problem that we have been trying to solve is this massive coordination. Amazon started from a monolith to a microservices architecture. That's where we clearly saw simple workflow, which was our first attempt at solving that problem. It was a game changer. essentially where developers can focus more on writing of their business logic rather than building that reliability backbone. I built that again at Microsoft Durable Task Framework, continued the journey. And then both me and my co-founder Max got reunited at Uber, where we built Cadence as an open source platform, where Uber was heavily invested into microservices.

12:38So you can imagine building stateful applications requires coordination across dozens of those microservices. So each and every iteration kind of build our conviction that this problem is not specific to places like just Amazon or Microsoft or Uber. Every modern organization is exactly feeling with similar challenges. What has changed in the post, like after the AI era, is the scale. What we have been seeing is the volume of use cases and the scales coming in as these AI systems is taking on more and more actions in the outside world. And the need of a system like Temporal to kind of be a reliable substrate for coordination and state management for these stateful applications so they don't lose what they are working on.

13:32And then you can provide continuation for the actual task that they are working on, generate the right outcome. The scale is massive. So if something like temporal was needed before AI, post AI, it's like a core requirement at this point. I think one interesting thing as well on this is like in the venture world, you talk a lot about founder problem fit. And in no other company I worked with have kind of the two founders been thinking about this problem for the last 20 years and really thinking about like, what is the system that nails this and gets it right? Which I think is why Temporal is so special.

14:11I do want to get to the round. Given the big news, LightSuite is leading this and, you know, it's seven months after a$300 million Series D out of$5 billion. What has changed between these two rounds? What has changed product team-wise in the last seven months? Yeah, I'm happy to take that, which is our core thesis back when we first invested was, one, there'll be a number of growing AI workflows, both long and short. And these workflows are going to require an execution layer that enables them to be reliable and successfully interact with, you know, core business processes in the outside world.

14:45and what changed and evolved over that time. So one, I think the inference in AI market just keeps scaling faster than we expect. We consistently just keep underestimating how quickly that market is scaling. I think the second is we're seeing the rise of these long running agents, which are just driving a lot more work and consuming a lot more temporal actions. I think we've all now started to get to experience like the power of what these agents look like. And then the other thing that's been cool is that a lot of traditional companies are trying to leverage AI to automate their day-to-day processes.

15:21They're reaching for temporal. For example, we spoke to a company prior to investing that has a mandate to automate as many of these kind of rote human processes as possible. And the way they're doing that is basically AI plus temporal to automate everything from compliance to vendor onboarding to the SDLC process. And so, you know, we approach summer and we're lucky to be doubling down. But that's kind of some of the story that led to this round. Yeah, I think the only thing I would add there is, by the way, like Anushka and Lightspeed and Ravi has been amazing partners for us. And if you think about like when we started the company, the business model we've been trying to build is we have an open source platform and we like generating awareness is so critical for us.

16:10So that's why we spend a lot in generating this awareness for temporal as a new category or a new way of building applications. That awareness eventually trickles down into adoption where these developers starts adopting this technology for our broad use of use cases. Then the way we monetize it, we have a cloud service, closing a logo, and then winning their first workload. And then from one workload to another workload, to the next one, to the next one, then we kind of really surely get centralized as an infrastructure offering within those companies. And so we've been building this machine.

16:50and the first half of this year, we started to see it compound. And this is where Temporal Cloud, I think, is processing over 1.9 trillion actions in August, which is up more than 350 % year over year. And if you look at our open source installs, August closed at around 43 million open source installs, which is almost 134 % compared to a year ago. And we have now over 4 ,300 customers running on top of Temporal Cloud now, which is up 138 % since last year. So we saw this build so much conviction that that machine that we have been building and really started to compound. And that kind of doubled down interest from like Anushka and Dean.

17:44That's amazing. Those numbers are so impressive. I hear also revenue is up like 200 % year over year by your own numbers and paying customers up as well. Just congratulations on what you've built. Now, to move from this big series announcement to what's next, tell me a little bit more about where this funding will go. What's sort of the next big focus as you scale to this next step of business? I think we definitely, Dampol as a company, is durable execution is a core foundation. that we have been layering. And then for us, we are doubling down. I think a lot of this funding is going to be doubling down and making a durable execution more and more reliable, create more primitive so we can expand the reach of kind of applications developer can build on top of the platform.

18:36A really good example of that is we shipped serverless recently, where today, for instance, we are in the business of kind of people run, use Temporal Cloud as a state management, but they run their applications themselves. But we really want to up-level the experience where people can ship their applications and we take care of the entire thing from running their applications and managing the state for it. So this is exactly where Temporal is going to be investing into these edges and categories to kind of up-level the experience for for our users who are building those mission-critical applications.

19:13And I think maybe this is a more macro-zooming out question for both of you. I think, as you think about this idea that, you know, we might have a world where agents at different companies call each other, or my assistant talks to, like, an airline's assistant. I'm curious, you know, how you view that all shaping and what the world looks like when all of this actually works seamlessly, you know, of course, with Temporal supporting it. Yeah, I mean, one thing that's so exciting is you're starting to see this with some products where you'll have agents talk to other agents. And so, you know, if I want to go to dinner with my spouse, I want my agent and his agent to know our schedules, pick a time, find a restaurant we both like.

19:55and then now think about what if you're trying to book this with your boss like you're definitely going to want it done correctly you're going to want it to pick the right day you're not going to want it to mess up booking the reservation in the right system and what you'll need is you'll need both sides to have the same guarantees you'll need both sides to complete the job without someone babysitting it and one thing temporal has launched and i'll let summer talk about it more But Temporal does this really nicely already inside companies. And so one team can reliably and easily call another team's workflow durably through a product that they have called Nexus.

20:34And so, you know, it's not a huge step function to imagine how we can extend this across company boundaries and enable that. I think Temporal could be a very important core primitive for agent-to-agent communication. But, you know, let me pause there and hand it over to Summer to share more about that. yeah so by the way Nexus is a huge bet we've been investing into it for a couple of years already to a point where within an organization temporal has started to become API boundaries how teams talk to each other and everyone knows like if you are building a synchronous service how to expose a synchronous service everyone knows that but then suddenly if you have an airway service which kind of you call and but the response come back in three days you pick three engineers and ask them to how to design such a system i can assure you they will come back to you with three completely different solutions for the problem and this is exactly what nexus solves is temple is already in the business of running these long-lived applications and very quickly people want to share these applications or apis with other teams so we build this backbone rpc backbone for long lift API, which is also provide the same durable execution guarantees that you see when you are implementing your API through a Damp Forward workflow.

21:56And I think we can clearly, as Damp, Anushka, the very next thing that we'll be investing into is like, why limit it only within an organization? Can we start providing the right guardrails where an organization can come in and then and share this thing, these abstractions, with other organizations out there through Temporal Cloud. So that's definitely the next generation that we want to take Nexis to kind of slowly start building more solutions and then eventually layer a marketplace for people to call through Temporal Cloud. That's really exciting. I really appreciate both of you hopping on to break down this fundraise for us and the exciting news.

22:37To close, I think since both of you do have this very unique window into where things are heading, and it sounds like Nexus is something to watch, as you think about the world running more agents in production and the broader industry still figuring all this out, what do the both of you think is something that's not obvious right now or will feel obvious in three years? I think one thing we keep underestimating is how widespread this will become and how much they can do. One thing that's crazy at Lightspeed, we just keep underestimating how these companies can scale and what they can become. And it's just a lesson I need to keep reminding myself.

23:21One thing for me that I see again and again, which is really under, by the way, this has been throughout my entire professional career. I built my entire professional career building messaging systems. As long as it's working, no one cares. But the moment there is a problem, like then people realize how important those kind of backbone is for the actual, like this is brand reputation. And once these systems are down, it causes so much pain to the customers. And it's actually erodes the trust in whatever product, whatever brand you are trying to build. And I think the thing which is really underappreciated is people thinking about reliability as an afterthought.

24:05And what I would, I think what I expect to see is like this is where people, especially where these whole AI is headed. and the massive scale that we are talking about and the huge impact in each and every part of like the, like how people live their lives, essentially, this technology is going to cause. Like I think reliability will cannot be an afterthought. I really clearly see it shifting that people will be thinking about what's the reliability backbone you want to have in place before you build these AI systems on top of it. Yeah, that's super insightful. I wish we had so much more time and we'll definitely have to have you all on for a full sit down conversation to break down where the space is going.

24:53But just really appreciate both of you taking some time out of your day to share the very exciting news and truly congratulations, Samar, to you and the team on this incredible milestone. Yeah, thanks a lot for having us. It was a lot of fun. Awesome. Thanks so much, Claire.

25:12Thank you.