Workflow Orchestration, Evolved

Hi, everyone. I'm Alan Shimel, Editor-in-Chief, CEO founder here at Techstrong. " This is a series of podcasts that we are doing with our friends at Temporal, and it's very timely given where we are in the world today around AI and agentic AI.

And this is actually episode three of the podcast. If you haven't caught the first two episodes, I strongly encourage you to go back, check them out. tv website, OTT app, which is available on iPhone or, well, any of the iOS, Androids, Apple TV, Roku TV, Amazon TV.

Whatever screen you watch stuff on, you could catch this on there. Episode three today, though, is about Temporal as the foundation, not reactive to AI, right? Temporal is a company that's been around well before AI was kind of the mainstream word that it is today.

But nevertheless, in many ways, timing is everything, and it was in the right place at the right time to help ride this wave that we're on with AI and to help people with their AI solutions. I want to introduce you to two people today, both on our panel. First of all, I'm thrilled to have back on Preeti Somal.

I've had the pleasure of meeting Preeti in person, and she's here with us today. Preeti, if you don't mind, would you introduce yourself to the audience? Tell them maybe a little bit about you.

Yes, of course. Thanks, Alan. Hi, everyone.

My name is Preeti Somal, and I run what we're calling the build, operate, and secure organization over here at Temporal. I've been here for three years, and it's been a fantastic journey, and I'm really excited to be on this episode talking about Temporal. Absolutely.

Thank you. Joining Preeti and I is Tim Hughes. Tim, welcome, and maybe you can tell the audience a little bit about yourself.

Yeah, sounds great. I actually didn't know that, Preeti, this is-- You're a repeat guest, so very cool. As a first-timer, I'm Tim Hughes.

I'm our SVP of sales and customer success. Been at Temporal for coming up on five years now. I'm excited to be here.

Absolutely. Tim, thanks for coming on. Of course.

So guys, look, I don't think there's any debate that we're in the age of AI, right? It's top of mind for a lot of people. " Well, there are some, but they're few and far between.

And also, this past year, I would say we've really seen the real disembarkation, the transition from purely generative AI to agentics. Right? Everyone, I think, realizes the real power of agentic AI versus just getting questions answered and so forth in generative.

So, the issue for many of our audience and the organizations out here is no longer whether to build with AI, but how to build with AI, right? It's about, I call it AI scale, right? So it's the speed at which you could do things and the breadth of which you could do things.

So how do you do that reliability at that sort of AI scale without breaking under the pressure? Because systems that were built for human speed, human scale, we're finding this out the hard way in case after case after case. They break at AI scale.

One of the suggestions, and I think it's where Temporal comes from, is the notion of durable agents, right? So Preeti, when we talk about durable agents, how would you define that for the audience? Yeah, absolutely, Alan.

So essentially, durability is all about delivering the value that you set out to deliver without any issues, and that's the core of what durable execution does. That's what Temporal solves for, and essentially applying that same concept of durable execution to agents to make them durable. Of course, I'm sure you've had this experience as well.

You're in a chat process with an agent, browser crashes, and you have to start all over, and I absolutely hate that. That's exactly what durability solves for, where if you can recover from the crash, whatever had happened, and it's just a much nicer user experience, and also really critical as agents start doing real business things, right? You can't have money getting lost or a booking getting lost because the agent crashed.

But yet, that is sort of what we're seeing, right? And part of it, I'm hesitant to blame end users or end user organizations, but in the gold rush to adopt agentics, they're not really system thinking, if you will, about, well, what does it mean across my entire stack, across all of my policy, process, and people But Tim, I would imagine in your role, you unfortunately are seeing this firsthand, right? Yeah.

And hearing some unhappy endings. Yeah. Well, maybe a way to respond to this, especially for members of the audience who are not Temporal-pilled or aware yet.

Temporal is in the simplest form factor for me is developer tooling for modeling or building long-running business processes. And it's developer first, it's code first. And we think about these long-running business processes as workflows.

And AI and agents is operating at humongous scale, obviously, but the world already operates at humongous scale. So our first big kind of, I would say, keystone customer was Snap, the makers of Snapchat. And one of our first use cases was bringing every Snap story onto Temporal as a Temporal workflow.

And those are massive scale. When you think about their busiest day of the year, which is New Year's Eve, and everybody has their phone out, taking photos of the ball dropping or fireworks going off, you're talking about billions of events across the world happening in a pretty compressed time period. And, those were some of the original core use cases inside of traditional businesses that Temporal was built to handle, and deliver to the world, along with things like every Uber ride is going to be a Temporal workflow.

And so really what we've found is that long-running business processes, the things that you build with Temporal today to power what we call the traditional economy, all of those kind of requirements map really cleanly to durable agents. Or maybe a different way to think about durable agents is agents who actually carry out work on behalf of a business that really matters. And so it's not just like, obviously, generating a funny image matters in some way, shape, or form, but it's a lot different than, to a previous point, completing a support case on an insurance website effectively.

And so just wanted to put that out there for those who aren't aware. Scale and Temporal, I think are pretty synonymous and kind of these long-running agentic workflows map very cleanly actually to just classic business workflows. Yeah.

Yeah. And Alan, if I may add, I think the key point that Tim makes is worth repeating, which is in order to have a scalable, durable agent, the business process that that agent is powering needs to be scalable, durable as well. Yes.

And that's sort of the system thinking that you were talking about as well. Yes. Exactly.

Right, because we've seen... Look, I've been in security 25 plus years. Wow.

Right? So we see what's going on now with this mythos, and that there's a few models out here that can do this kind of vulnerability discovery. The systems that we've built for vulnerability management are just not built to scale to deal with the amount of vulnerabilities that these new tools are finding.

And so, that's why you see Microsoft putting out 500 plus fixes on Patch Tuesday, where it used to be in the 50 range. So it's like 10X what used to be normal. To me, that's a great example of AI scale.

Now, my question back to you is, how does durable agents, can durable agents make up for a system that's not designed with system thinking, or do they work hand in hand there? What's the relationship? It's a great question, and I think this is where the power of Temporal really shines, right?

And we see some of this even with some of the coding agents and associated GitHub set of issues, right? The core system behind that durable agent really needs to be able to scale durably as well. And being able to deliver that through a single platform like Temporal is where really the magic comes through.

Yeah, Alan, I think, and as a side note, I hadn't thought about Patch Tuesdays for, oh my God, so long. But you're absolutely right. The scale, the AI scale means that all of the systems need to be able to respond and live up to that scale.

And people feel like, okay, it's just the agent that needs to scale, but really AI is going to need us to scale every layer of that workflow, of that business process that powers the agent. I'll build on that too, because there's lots of reasons that people use Temporal. Obviously, what we're selling is reliability as our core value prop, but there's a lot of things that come along with the Temporal ecosystem.

And when it comes to things like scale, so let's say that agents are allowing you now to go off and do many more tasks all in parallel and do more faster. When issues arise, it's still eventually going to hit a human to look at, okay, what is the issue, what's the information, and then what are we going to do to course correct this thing that's arose? And Temporal really shines in that scenario as well because of the way that the system works and how we provide this durable kind of outcome is all about things like our event workflow history.

We can actually systematically see what is every step that happened in a particular process, and then what broke and how can we actually triage that. And so Temporal pulls together a number of really nice ideas and concepts all into a single product that maps again into this exact type of scenario. Okay, we're doing way more at way bigger scale, but we're also doing it in a reliable fashion, and part of that reliability story is visibility into the execution of any given workflow, which then allows you to make appropriate business decisions when things go right or wrong, but you need to investigate it.

So what I find interesting is, look, as I think I said at the outset, Temporal's been around a long time, way before LLMs were a glean in our eye, a glint in our eye. Right? And so it's easy to miss, I think that Temporal, what it was built for, what its mission is, was never built, perhaps with AI in mind, certainly not exclusively for AI.

Right? There was a Temporal long before we were thinking about generative and agentic and LLM kinds of things. You've been powering mission-critical workflows right across the key verticals, finance, healthcare, e-commerce, et cetera.

So one could say, Temporal was fine before there was an AI. Mm-hmm. But when you look at it, I guess in history it's going to go down to 2023 when ChatGPT, that release, I think it was 3-0 or whatever, I think November 20.

That's going to be sort of the defining era of the beginning of the AI era, right? Since that time, what's changed from pre-AI Temporal to post-AI Temporal, or is it the same Temporal? Look, I think there's lots of things that changed, and I think a lot of different...

So for the audience, I don't have a technical background. I sell technical products, so certainly, in getting into how it's impacted engineering, I'm going to defer to Preeti, in a lot of different ways. But ultimately there's a ton of commonality between the two worlds.

And now I think we are just thinking more about how our product impacts the business from multiple different lenses. So I'll give you one kind of narrow example would be is, developers are... Some of the core value props of Temporal is the developer SDK.

It is like the ergonomics of actually building with our product. We have these fantastic APIs or primitives that create a really nice abstraction around really hard concepts that are tough to reason through in the absence of something like Temporal. So what is that going to do to our product and how people use it and engage with it?

It's definitely going to shift it over time as more and more tools are just like, or maybe more products are built through code gen. But ultimately, the core value props of what our system delivers and what we're doing, that stays the same. An LLM is now just a part of a business process, and those business processes continue to be the same, and the characteristics that a business requires out of those processes, those are also the same.

It needs to be up, it needs to be reliable. When it goes down, it needs to be able to recover effectively and all those types of things. And so I guess from my point would be is nothing's changed, just like kind of the aspects of the playing field or the dynamics of what we're dealing with, have shifted slightly, but the core underlying baseline still the exact same, and the value prop that you come to Temporal for is reliability at scale, and that I don't see a world where that goes away in any way, shape, or form.

Preeti, Tim kind of said off the outset, right? He's not here giving me the technical aspect of it. He said you will.

So I'm going to ask you, from the technical point of view, what do you think? So, I think what Tim was sort of alluding to from a technical point of view, the way that the Temporal applications get written is starting to change, right? So, we definitely are a developer tool, the developer productivity and reliability kind of being two sides of the coin here.

And we know, of course, with the AI coding agents, the developer productivity side has exponentially gotten faster. And so the way we're responding there is making sure we've got all the right skills and plug-ins to really help developers code Temporal using all of the coding agents. But the core of what we solve for, the core innovation, the programming model paradigm around a workflow calling distinct services where anything can go wrong, that has only gotten more focus.

It's sort of we were built to meet this moment where not only has it gotten more focus, the challenges around LLM responses, consistency, cost, all of these challenges that AI have sort of brought to the fore are challenges that Temporal already solves for. And so, the core technology, yes, one thing I will say from an engineering point of view, the scale of what we had imagined it's hockey sticking a lot faster, right? So we were pre-AI, pre that defining moment you were mentioning, Alan, we were sort of growing really nicely.

But the scale of the kind of the transactions we process on the Temporal Cloud platform, you can clearly map the growth of that with how AI is starting to get adopted. And so I think that's the biggest thing for us from an engineering point of view, is we are hitting scale like 10X, 100X sooner than we thought we would. Absolutely.

I'll throw out something else on this topic that I've been thinking, that keeps on coming up in my mind all the time, or maybe not in my mind all the time, but comes up quite a bit, which is like, I think this kind of idea that LLMs are just become the entire thing. And Andrej Karpathy had a really good tweet maybe a month or two ago where he was talking about you don't use an LLM to reason about a retry policy. I don't need an LLM brute force guessing what is the appropriate retry policy all the time.

That is a thing that you "hard code," but you define a retry policy within your application, and you just follow these rules over and over again, as opposed to shoving an LLM to that part of the process. So that's one thing that I would just throw out there that definitely comes up for me quite a bit, is reasoning about, well, when is it a good time to pull on this tool, and when is it a good time to use something like Temporal? And that is a fantastic example.

You're implementing a retry policy in prod. You are not asking Claude to try and figure that out on the fly every time. You're 100% correct, Tim, and it's a symptom of the stage we're at in AI maturation and adoption is we're just using AI for everything.

" Right? Between the token cost and everything else, I'm better off having a person do it. Yeah.

I'm better off. I don't need to take a sledgehammer to kill every fly. Yep.

That's what we got fly swatters for. I do think this is a stage that we'll grow through as we learn what makes sense and what doesn't make sense. Oh.

But it also brings up a bigger issue, and that's orchestration. Right? Yeah.

I've written that the AI stack is the cloud-native stack, for instance, right? Because they're using that Kubernetes-based orchestration and the whole everything that goes with it to help us deploy our AI applications, so to speak. Mm-hmm.

At the same time, platform engineering, another key audience here at Temporal, are understanding that we need the AI built in. We need AI to be built into the platform in order to manage AI with the platform. Mm-hmm.

It's the same, I think this is a key piece of it, and it's a key piece of what Temporal does. We need to manage and scale, whether you hear talk, 10X the amount of people. We're going to have 10X agents per person, 100X agents per person.

You need to be able to manage and scale that the same way we manage and scale our platforms, the same way we manage and scale the stack that these things run on. Temporal manages and scales agentic workflow. Yeah.

And that's now, that doesn't exist in a vacuum up here. That's built in. It has to be built into the platform.

Yeah. It has to be built into the stack. Yeah.

Right? It doesn't exist in isolation, so to speak. Yeah.

People want this holistic kind of management tool, if you will, or how do they operate? It's an operational- Yeah, I think that is-- I hear you loud and clear on that. It's interesting, right?

So LLMs come on the scene, code gen starts coming on the scene, and I think that we're very engaged with folks at that point in time, say, I don't know, 18 months ago, but were they using Temporal in earnest? No. And why would I say that?

It was a lot of experiments. It was like, "Hey, we'll plow tons of money into this. We will try anything and everything," and our expectations aren't that you're going to have some magical outcome that transforms the business.

I think the expectation was they were going to learn and identify places that they can then actually deploy this technology effectively. And then where we really saw, I think, a big takeoff over, let's call it the last four quarters or so, has been large organizations getting serious about moving these things into production, having concern about production concerns. And production concerns are the things that you're talking about.

And Temporal is specifically built for that. It is the platform where you can put your agent, which we'll largely think about as a long-running workflow, which long-running can be three seconds or 30 days. And we're going to provide you the characteristics that you care about in a variety of scenarios that production applications have to care about.

Things like, okay, you call a third-party API, it doesn't respond. What do you do? You have a long-running agent that's going to be calling lots of tools and creating context and all these different types of things.

When it crashes, you certainly don't want to start that process over again. You want to rehydrate and pick up from where you were. So basically, my point would be here, as we've seen organizations pivot to, we need to actually start seeing ROI on this then, which means we need to be in production, which we need production-grade capabilities.

That's where Temporal gets pulled on, and that's why Temporal, I think, is fun in this space in a lot of different ways, is we don't have to go out to retrain an entire workforce on, okay, now you got to learn all the ins and outs of what it is to be this LLM thing and all of these really nuanced kind of concepts that it's very hard to be up and running on. We just have to continue to focus on what does an engineering team that's delivering production-grade applications need that Temporal can do, and it's very similar if you're running payment rails on us or if you're running an agentic workflow on us, a lot of crossover, and that's, you might say, my distillation of the engagement with enterprises over the last at least two quarters. Yeah, absolutely.

I think a couple more things I'll add on. Governance is a big deal, and the visibility that Temporal provides, that Tim was talking about, gives you that element of governance. So you can basically see all the steps that your agent is, or your AI, whatever application you're using, has performed.

You can export this to an audit log. You can kind of get those capabilities with Temporal. And I think the second one, and I feel like the industry doesn't talk about this much with AI, but it's the concept of open.

As you know, Alan, Temporal, at its core, is an open source project, and the way that we see enterprises making these decisions around ... the management and control plane for agents, this is important to them. They want the ability to have an open platform that can plug into whatever other systems they're using, whether it's their observability stack.

This is literally everything Tim was talking about, where that set of requirements is not new for AI. That set of requirements has existed with platform teams for any application that they currently run. It's something we understand deeply and deliver against already, and that's where we're finding a lot of the adoption within the enterprises really accelerating with an extension of what they're already getting from us.

Guys, I want to return to the example of the AI stack again, to the cloud-native stack again. Right? One of the things we saw with cloud-native adoption was if you were building from a greenfield, from a fresh sheet of paper, yeah, you could go with Kubernetes and containers and microservices, and all that that entails, and you had a really modern application that was scalable and good for you.

Unfortunately, the world's not made up of greenfields, right? The world's made up of brown, muddy, messy installations of platforms and things that have kind of been built organically, piled on top of each other. And so you go into a company and their finance system isn't some shiny, new AI-enabled cloud native system.

It's waterfall. It's a mess. Tim, how does Temporal help that, right?

How do we go in there and say, "Look, you can't throw this all out. " How do you do that? Well, first and foremost, that's why AI's been also exciting on a company building perspective, right?

Greenfield is every AI application is essentially greenfield, and so I think that's where you also see people move really fast. And so it's an exciting area to invest in and partner with folks on. But what you're describing is the normal engagement model for Temporal.

"Hey, we have this monolithic stack. It's 20 years old. It powers our business.

It powers the most important part of our business from a revenue generating perspective," blah, blah, blah. And we almost always start with our customers on something like that, and where our conversations start are, what does this thing do? How does it work today?

Okay, how does it actually drive your business? And then show us the areas where it becomes a problem for you. Where is your SRE nightmare?

And oftentimes, there'll be a part of that monolithic stack that is super problematic that we can identify and wrap in Temporal, and we start with just a small piece of a brownfield application and help you triage or course correct the thing that's your heart, your biggest issue. " And that's an extremely common pattern for us, and it's honestly one of, I think, the superpowers of Temporal. You can start incredibly small or incredibly big, and we're going to add value kind of right from the jump.

But I'll pass to Preeti. I think she'll probably have more of an actual practitioner's lens on it, but my point would be is that is our starting point, and we always just start small and grow from there. Absolutely, Tim.

I think not much more to add beyond sharing that some of the terms I've used customers call this pattern is hollowing the core. So they'll kind of pick a piece of that monolith, the messy stuff, and it's often triggered by either business reasons or honestly, some of the legacy vendors that have maybe outlived their relevance that the enterprise is finally needing to kind of move off of. And the really interesting thing that's happening, Alan, we're seeing with the coding agents, this transition from a legacy vendor to a temporal workflow can actually be sped up quite a bit.

So even that process of modernizing and hollowing, AI is actually helping customers with that too. Guys, we're running low on time. I want to kind of tie a bow on this in here now, right?

I think what we've seen here today is really that Temporal is a tool that's been around, that people have used, counted on, relied upon, critical infrastructure built around it and with it, meeting a little bit of Muhammad comes to the mountain. Meeting this new object that's changing everything, right, and how we do it, but yet Temporal has proven itself to be a good tool, a good platform, a good way of managing this latest and greatest agentic AI kind of wave that we're seeing. " Right?

Tim, you got to sell that, though. Mm-hmm. Right?

How do you, when you sell it to someone with that, "We're built to last," give us the background. How do you build up to that? How do you make that real?

I don't know if I have a great answer, just a clean response to you, but I will speak to a few different things just about how to think about Temporal if somebody in the audience is interested in it. First and foremost, we recognize that the things that you build on Temporal are mission-critical applications for your business, and Temporal is an MIT open source piece of software. We're committed to that.

It's a big part of how we go to market and how we actually give companies and organizations the confidence and trust to partner with us long term, because we do provide a very clean off-ramp to go do something else. The next bit is that we are a consumption-based platform. We earn every workload.

We earn every single application that runs on us. And we also expect to start small and have to earn your trust over time to actually deliver those outcomes. And I think that's what you're going to expect if you start to actually engage with our product or our organization in general, is it's hyper-focused on how can we actually drive an outcome and how can we actually become an extension of your team to get these production applications to run reliably at scale.

So I don't think I have anything magic to say beyond that other than I really encourage people to try it out. I really encourage people to engage with our community and our company directly. I think you're going to find a large number of very competent people, very focused on what is your issue, what is your problem, how do we get past it, and how do we get you back into production.

I love it. Preeti, I'm going to give you the last word. How do you keep Temporal relevant, current, on top of what seems to be an ever-changing world?

Absolutely, Alan. I'd say two things. First is the depth of the innovation that Temporal provides.

The fact that it spans these technology waves and changes really speaks to the value that lived experiences from the co-founders resulted in building this programming model, right? And so the fact that we can run a Snap story and financial transactions and AI agents, that really speaks to how deep the thought process and the innovation here is. And I think the second thing, it ties to what Tim was saying, we are deeply, deeply rooted in the success of our customers.

We are with them. And because we are with them, we are providing that stable foundation. We are really getting a ringside seat view into how enterprises are thinking about AI, and we're actually shaping that.

More and more of our customers are asking us to help them grapple with how they should be using AI, and that's how we're keeping up with this. Nothing, in my experience, ever beats the boots on the ground, in the trenches with customers, learning and then shaping product direction based on that. Preeti, Tim, thanks for coming on to today's edition of the podcast.

We appreciate it. Thank you all for watching, listening, whatever you're doing, whether you're watching or listening this. As I said, this is episode three.

We'll have episode four soon. Go back, check out episode one and two. But until next time, this is Alan Shimel.

Have a great day.