From Prompt to Production: Building FlowAgents That Use the Tools You Already Have

Karan Munalingal • 00:04

Hello, everyone. Welcome to another Itential webinar. In today’s webinar, we’re going to be talking about how we take agents from prompt to production. So, it’s going to be all about building flow agents, leveraging Itential’s FlowAI, Agentic Harness that use your existing tools that you already have today and put them to work on your behalf. So, I’m joined by my friend Joksan Flores. Joksan, would love for you to introduce yourself and then we can take it from there.

Joksan Flores • 00:36

Hey, Karan, and hi, everybody watching. My name is Joksan Flores. I’m a principal SE at Itential, and I work with Karan with all of our customers on GoToMarket and nowadays on how do they take agents into their environments and modernize the way they operate infrastructure.

Karan Munalingal • 00:54

Thanks, Joksan. And again, hello, everyone. Karan Munalingal. Head of the AI Strategy and Innovation for Itential. So let’s get started. What is a digital coworker? You know, working with a lot of our customers and hearing a lot of our prospects in the field, you know, based on their journey into the AI ecosystem, a lot of folks already have exposure to things like chat GPT, co-pilot, et cetera, right?

Karan Munalingal • 01:24

And that particular exposure focuses directly on you being able to ask. LLM questions and responses back, right? So it’s a lot of QA going over there. What we’re going to talk about today is how do we take the next step in enabling our customers and the industry to essentially build yourself a digital coworker, right? Something that takes advantages of your instruction, your knowledge base, some of the skills, the tooling that you can provide it, and basically act on your behalf like you would, right? And this is, these are the types of agents that we called React or goal-based agents that not only respond to your questions, but also will take actions on your behalf more like in an orchestrated way, just like you would. Imagine me being super busy, but I need Jackson’s help to actually do some of the work that I need done.

Karan Munalingal • 02:22

I would basically write instructions, provide him with the tool sets, as well as some of the essential skills that he has or I have. And then he will go actually go take action and do that in certain environments for me. I don’t know, Joksan, if you want to say your piece on this, but this has been a very interesting conversation with a lot of our existing customers who have had early exposure to AI technologies in general.

Joksan Flores • 02:49

I think this gives people a way to leap forward on the utilization of AI, right? I think traditional AI, just generative AI using chatbots is super useful, but it’s also very much a one-time thing. You go through the operation, you did the thing that you were going to do, and then you’re done. These agents are meant to be design, time, built. And then from a service exposure perspective, you’re exposing this as you would any other service, right? Any other workflow on our platform or code through any other platform, right? It’s still a service.

Joksan Flores • 03:19

But now reasoning providers the capability to doing this very, very, very quickly. And we in the attention platform provide all the tooling, all the harnessing, all the RBAC, all the controls that we need to make that use case work in a very flexible way. And it just takes advantage of the L of M’s reasoning power and takes it yet to another level.

Karan Munalingal • 03:39

Thanks, Joksan. That’s a great segue to, you know, Joksan mentioned a harness. So what does it take for someone to actually build these digital coworkers or React or goal-based agents, right? So Itential is, you know, has already brought its own agentic harness for all of our customers and future customers that now have an opportunity to build these digital coworkers that could be, you know, part of their team to support some of the business objectives that you have. I continuously keep hearing from executives that they want to do more with less, right? So now Itentual is providing an agent harness where you can actually bring your SMEs and their knowledge to the table, but they’re essentially building a coworker that will help them do their mundane activities so they can rather focus on some of the more critical objectives that their executives want them to, right? So specifically in this example, we’re going to talk about two things.

Karan Munalingal • 04:38

One of them is a flow agent builder and the components of the flow agent builder. And we’ll talk a little bit about the deterministic tools. Right. There’s when you’re building these React or goal-based agents, you need to provide it context, right? And context is a multifaceted thing. One of them is a set of instructions. So you’ll notice here, these are prompts, right?

Karan Munalingal • 05:03

So you basically need to tell your agent, hey, these are the instructions I want you to follow. Do this, do not do this. If you see this, do that. And if you run into this error, exit or move forward or introduce a human in the loop task, right? So these are some of the instructions that you would want your agent to follow. The 2nd thing, as part of the context, setting context for these agents, is a tooling. Essentially, when you ask, like if I were to go ask,

Karan Munalingal • 05:31

Jockson to go do something on my behalf. He’s going to ask, well, how do I do it? Right. I’ll give him the instruction. He’s like, well, what material do I have to finish the task that you asked me to? So these are the tools. So for an agent within the infrastructure, you know, you want it to have access to APIs, human-in-the-loop approval tasks, potentially your scripts that folks have already written today, MCP servers and MCP tools that are made available by a lot of vendors today.

Karan Munalingal • 06:02

Maybe existing workflows, a pipeline or another orchestrator that you’re currently using that has a very deterministic flow that takes an action, right? So think about everything that a human does today, essentially by hand or by writing a script. All of your existing investments essentially become tools within the flow AI agent builder experience, right? In this particular example, you’ll notice an organization has three to four different teams that are focused on certain technologies, right? So whether it’s a DNS or a DDI-focused team, they can now build an agent specifically by just using APIs provided by Infoblox or ROT53 or BlueCat, right? And then they might say, in order for me to build this agent, I might have to have human in the loop. So that’s another tool that they can associate, right?

Karan Munalingal • 06:55

But then now you go to another team that’s focused on tier one, tier two operations, where they’re essentially either running commands or scripts or leveraging, you know, PyATS MCP server to collect data from the network and then pushing that into ServiceNow tickets from a data gathering and RCA standpoint. Right. And on and on you go. So you’ll notice as different teams get onboarded within a platform that provides a harness, they might have a different set of requirements in which tools they already have available or the ones they might want to bring on later on and then use. Right. So that’s essentially tools are the 2nd part of the context. And finally, it’s LLMs.

Karan Munalingal • 07:40

Right. This is where the agent goes and reasons through your instructions and picks which tools to utilize in order to do the job that you asked it to. So this kind of formulates and Flow Agent Builder enables you to define all three to make sure that your agents are purposefully doing things in production that you asked it to with the right tools and with the right level of intelligence on the reasoning. Right. So that’s a quick little blurb on the harness itself. You will notice one thing up top. I always mention this.

Karan Munalingal • 08:13

It’s great that people have the ability to build things faster, better, easier. But if no one is consuming it, it’s useless. So in this case, it is very critical from a platform standpoint that, you know, Atention is bringing the ability, the harness for you to build the agents, but we also have the native capability to expose these agents as a service, not only for other agents, you know, in some of the other systems that customers are already using, but existing pipelines, humans that are logging into ServiceNow and requesting services, right? It all matters is how do you meet them where they are so they can consume some of these agentic services that you’re building. So, the next piece is huge, right? When it comes to automation orchestration, we have seen a journey where a lot of our customers and the market and industry in general have taken where they’ve introduced a lot of guardrail security governance before automation can do anything within the environment or the infrastructure. Now, the conversations that we have had recently with a lot of executives, there is even more focus on the platform and the non-functional capabilities before taking the agents into production.

Karan Munalingal • 09:31

And some of them that are called out here are, you know, access control, logging, auditing, insights on what the agent is doing, access to integration, you know, SSO, like with respect to. Getting access to building on the builder side, but also getting access to consuming the agent capabilities, right? So these essentially make up the foundation, even though they’re non-functional requirements and capabilities. Essentially, for an agent to effectively execute securely, you now need all of these around it in order for it to actually do a good job effectively. All right. So talking about the platform in general, right? So talking about taking agents into production.

Karan Munalingal • 10:19

What matters more is the foundation that you built it on. So, for the longest time, we’ve had a lot of customers, you know, going on their automation orchestration journey, and they put a lot of effort into building a moat around: hey, do I have good access control? Do I have good logging? Do I have good auditing? Because if I’m going to let automation change infrastructure, I want to have full control and needs to do that securely and have an audit trail all the time. So, when we’re talking about agents going into production, now the requirements are even more robust, right? So, if you look at some of the capabilities that were now an impending requirement for automation, now folks are looking at it as a must-have before an agent touches the infrastructure in production.

Karan Munalingal • 11:08

Things like being able to understand what the agent is doing from an auditing and the logging side, who has access control to build the agent, but also execute the agent. Agent having access to infrastructure with respect to leveraging secrets management, providing insights on how effectively it’s doing its job, how consistently it’s doing its job, right? So, this is the bigger part. And within the itential ecosystem, right? The platform in itself. Has provided this capability for our customers that are automating and orchestrating, but now also for agents naturally, because Flow AI sits within the platform. The 2nd big piece is execution at scale, right?

Karan Munalingal • 11:51

When you’re writing automation, it’s great. Now you’re running it against your infrastructure. But when agents are doing it, they’re thinking, they’re invoking tools, and they’re doing it more rapidly, right? So, what also matters now is a foundational unified execution engine that can support many different strategies on how you can change the infrastructure. APIs, CLIs, your existing scripts, your playbooks. This is what we’re talking about when we’re talking about leveraging your existing investment and then layering AI on top and taking that into production, right? So, this is a bigger part and big discussion that we’re going to have when Johnson also talks about the demonstration piece.

Karan Munalingal • 12:36

And he will highlight some of the pieces that have become more important for our customers than just being able to build these agents. So, the last piece, right? Agents don’t replace what you have built. They actually leverage. So, the more stuff that you have already had, that’s great. But, you know, in addition to what you’ve already built, there are existing vendors that y’all are partnering up with that are providing some level of automation, some level of APIs that can do things more effectively. And now, as part of the next phase, MCPs have become a thing, right?

Karan Munalingal • 13:11

So, a lot of vendors are now providing MCP servers that provide more inbuilt capability that you can now leverage right at the gate to drive faster time to value. And when you’re building some of these agents, they’re not one-off agents. They become reusable by default because you are leveraging the existing APIs, the existing scripts that you’ve been running for years and improvising in your existing automation strategy now just becomes agentic. The final two pieces are built-in governance. It becomes very important that governance is always. At play here when you’re building agents and taking them into production, because just like your automation journey, your agentic journey is going to take the same path until you become more assured that the agents are doing the right thing, right? So the governance, security becomes a big thing and simpler to build.

Karan Munalingal • 14:07

Finally, this is how do we have an opportunity to bring SMEs to the table that may and may not have coding capabilities, but they are the experts in specific domains in the network. So how do we enable them to start building more to support more innovation as well as results? So, with that, I would love to hand it off to Jockson so he can show us some of the capabilities within the platform and how some of our customers are taking advantage of it to drive agents and the agentic operation mindset into production.

Joksan Flores • 14:43

And yeah, thanks, Karan. I think one of the things we’re going to be focusing on building is going to be something very basic that’s, well, very, you know, with some depth to it, but at the same time, something that is based on stuff that I already had laying around, right? We’re going to be leveraging a couple of scripts, an API into a system that we have, some of the human in the loop activities or tools that the platform provides. And if there’s a theme that I want people to get out of this is you can build fast, but at the same time with confidence, right? You talked about all the governance and all the aspects that we provide, right? I’ve gone the whole route of building all this stuff by myself with a custom UI and everything, you know, using chainlit UI and Lang Chain and all that stuff. I wouldn’t put that into production, right?

Joksan Flores • 15:26

It’s functional, but it doesn’t do anything else for me, right? It doesn’t have the integration, the protection bits, and all the stuff. So, if there’s anything that I want people to get out of it, we can build quickly with confidence. We can choose to do read-only or not activities at 1st , right? I think a lot of executives and organizations are very bullish on AI technology. We’ve seen that, but there’s also been like, okay, we’re going to go ahead and do something, but we want to make sure that we have confidence how we’re acting into the environment. And we’re trying to provide people a platform that provides them with that confidence for them to build and be able to execute and have all the governance and all the other aspects that they need, right?

Joksan Flores • 16:00

All the boring stuff that they need. Okay, so we are in the attention platform in the agent projects section, and we have a lot of stuff going on in here, but we’re actually going to start from scratch. I did bring my own prompt because I am fairly lazy and I don’t want to type the whole thing. But there’s a few things that we’re going to do, and we’ll go and explain this and take our time kind of walking through all the bits and explaining how this works. And the 1st thing that we do here is we create a project. Now, projects here, you’ll see there’s a lot of other stuff going on. We’re going to go and create one that’s a new one.

Joksan Flores • 16:30

We’re going to do some application infra pervisioning. I’m going to call it that. I don’t think I’m overlapping with anything, and I’m not going to put a description at the moment. So I created a project, and a project is a container of agents, right? And typically, what we do is we, you know, there are several ways to go about this. We typically make them surrounding a particular use case. There might be certain use cases that might need several agents because you’re doing crowd activities, right?

Joksan Flores • 16:57

We have a, I think we have an F5 project here that will do crowd activities on load balancer VIPs and so forth. And we have an agent that does creation, an agent that does modification, and an agent that does deletion, right? And we, you know, use other platform bits to keep and gather state and so forth. But essentially, the whole idea is that there’s a project where we can self-contain some of these agents and the stuff that they do in some sort of way that makes sense for the organization. The other thing that I will highlight here really quickly is on the project settings section. This project has been created for me and by me, and I have not exposed it to anybody else, right? The reason why, because if I’m doing testing and I’m building at the moment, I don’t want anybody else to have access.

Joksan Flores • 17:40

Now, Karan is here with me, so I will actually go and add him here. And if you look at there’s several things going on in here, and we have group permissions based on RBAC format and so on. And then all the user lists that you saw there come from our IDP integration. Right. And kind of, you know, going a little bit over the boring bits. And the reason for that is because we want to talk about that confidence, right? That platform that enables this kind of capability inside of the platform car.

Joksan Flores • 18:06

And I think, you know, we talk to a lot of customers. This is super important to them. You were highlighting this.

Karan Munalingal • 18:10

Yeah, 100%. I think the very 1st question that every team leader asks is, how do I keep from agent just propagating it all over the place? Right. Because they have had that experience with scripts. When someone puts it in a centralized location, someone gets access to it. And now it becomes everybody’s scripts when it was not ready. Right.

Karan Munalingal • 18:29

So essentially with respect to when you’re looking around just. Software development lifecycle. I know there’s something out there around agentic development lifecycle. So the same paradigm applies, just like when you’re writing a piece of code that until you’re ready, you’re not going to share it with your peer or anybody else in the organization. We’re trying to bring the same level of governance for the builders of ours who are the SMEs, the experts. We don’t want them to have to worry about, hey, is it going to get leaked out, et cetera? So they have full control of who do they share the agents with.

Karan Munalingal • 19:05

So, someone can test it and find on it into production, or someone can reuse it as part of their own agent orchestration.

Joksan Flores • 19:12

Yeah. And just to highlight another piece, Karan, that you touched on that I didn’t highlight, it’s you could do grouping, right? So, there might be several teams, you mentioned this during the slide before. There might be several teams leveraging the platform. There might be a networking team, a DNS team, a security team that are building. And guess what? We’re trying to make it super easy for people to build.

Joksan Flores • 19:30

But one side effect, which is a good side effect, is the more that you make it for people to build, the more that they build, right? So, you saw we have a plethora of use cases here, and it’s not going to stop. We’re going to be building lots of lots of these agents. So, we want to make sure that we only propagate into the people that need them and we provide them access to execute them and so forth. And there’s several levels. So, we’re going to go and show, and there’s actually another level inside of it where we’re going to see a few other things. So, let’s go ahead and start agent creation.

Joksan Flores • 19:57

And we’re in the agent builder section. Karan talked about agent builder. We also have the sessions, auditing, and so forth. We’re going to be building 1st , and then we’re going to be exposing, and then we’re going to be auditing the agent execution and doing human-in-the-loop things. So, for now, let’s go into building and let’s do AWS application info provisioning. That’s going to be our use case today. Make sure that I type properly.

Joksan Flores • 20:19

Yeah, I can type. Okay, cool. So, we got our name of our agent here. Like I said, I’m going to bring my prompt, but we’re going to go walk through it piece by piece. And I have my own idiosyncrasies of how I create these prompts. And I like them in Markdown because they render quite easily. But essentially, the idea here is we are going to provision an AWS VPC and an EC2 instance inside of it.

Joksan Flores • 20:41

Before that, right, before having the ability to do agents, my paradigm was either I have a couple strips, a couple scripts, right? A couple Python scripts that I use the Bodo3 library, right? And I just run them manually from my environment, right? That’s kind of the idea, right? And I’ve heard this from many, many, many people where they just say I have scripts that work, they’re pretty resilient. I’ve been using them for a long time, and they just create things from my laptop and I pass them environment variables and so forth. So, what I have done is I have onboarded those scripts into our gateway.

Joksan Flores • 21:13

They have secrets injection to them using environment variables the same way that I have deployed it on my laptop, but I’m actually using the gateway secrets management function in order to pass in the variables that it needs. In this case, right, the AWS secrets and so forth. And my scripts, they need a few things, right? They need a few parameters and so forth, right? They need, they’re going to be creating a VPC, so they need a VPC name, they need an instance name, and so on. But in this case, I have actually decided to simplify within the prompt itself. How do we do some of these things?

Joksan Flores • 21:46

So I have laid out the inputs here at the beginning, and we’re using, you know, double mouse dashes. So Jinja syntax in our flow AI platform. That’s what we have opted for to provide for our customers. And you can see there are some errors because I haven’t declared the variables yet, but we’ll do that in a minute. So I have all my variables declared. Now I have simplified all my inputs, right? I have said I need my agent to worry about having collected app name, region, template, requester.

Joksan Flores • 22:11

Now the platform will make sure, and again, this is the error down here. The platform will make sure that those variables are supplied. And they will be useful in a couple spots, right? Once we do it here during build time, but also during exposure time. This is what my users will need to provide in order to leverage this agent. I also have a policy that says the name pattern for both VPC and instance is the app name, right? Now, I could make this app name colon something else, right?

Joksan Flores • 22:37

I can enforce my standards. And this is how I tell my agent, hey, this is how you’re going to do things. In order to keep things simple, I have a look and done some of the fixed facts of some of the network containers and so forth that we’re going to use for InfoBlocks to do a network allocation for us. So in this case, it’s all fixed. Again, this could be something that you provide variable through a drop-down, through some other mechanism, and/or even a human-in-the-loop activity that this comes from somewhere else, from a ticket, and so on and so forth. Lots of things you could do. And then we get to our step down here.

Joksan Flores • 23:08

So the 1st thing that our agent’s going to do is confirm that region is supported. My scripts have a method of exposing their Their parameters that they take in, right? Their positional arguments, and those parameters will serve to expose what regions we support. Then, also a template, right? Templates, the same thing. They are supported.

Joksan Flores • 23:30

I’m going to pass those in, but we can put them in drop-downs in the platform and make them available in different ways. Then, the 2nd step is: I want an approval, right? And this is our 1st human-in-thloop activity, right? Karan, we’re keeping this super, super, super, making sure that very governed, especially because we’re doing resource creation. We want to make sure that our customers have the ability to control, right? But notice that I’m just saying approval number one, present the request, ask for approval to proceed on rejection, stop, report, reject it.

Joksan Flores • 23:58

All natural language skills, all natural language typing here. And I’m just telling my agent, hey, this is what you got to do. This is what I would tell a tier one engineer when I’m building a mob, right? So, rather than me as an architect coming and building mobs document in a Word document or in a confluence page. I am building it directly here into the prompt. Then the agent will allocate a subnet, create the VPC itself. Well, giving it, we’re going to give it a tool for doing that.

Joksan Flores • 24:25

It’s going to launch an EC2 instance, then it’s going to confirm the deployment, and then it’s going to provide a report. I’ve also done, you know, kind of gone a little bit over the over the top here and providing an attention branding instructions. These instructions could come from multiple places. I have chosen to keep it simple for now. So I kind of want some colors to be attentive branded, like backgrounds, navy text, and so forth. Now let’s go back and do the variables, right? We talked about adding variables into the prompt.

Joksan Flores • 24:50

We’re going to do app name is already rendered here. So we’re going to just add, add, add, add. I’m going to click on these. I’m going to validate the data types, app name, template, request or regions. They’re all strings by default here. That looks good to me. I am totally okay with that.

Joksan Flores • 25:05

That’s kind of what I need. And then we’re going to do tool selection. So, we’re going to do a few things. The 1st thing that we’re going to do, our agent’s going to do is it’s going to do approval human-in-the-loop. So, we’re going to use view data. That’s the task that we’re going to use for our human-in-the-loop interactions. Then, we’re going to do assign next network.

Joksan Flores • 25:26

And because we’re using InfoBlox to allocate networks, this is the tool that I need to use for that. And this is the API that InfoBlox will let us use to pass in that network reference that we provided in the prompt earlier to allocate the subject that we’re requesting, right? We’re requesting, I think, a slash 28 in the prompt. Yep, right here. We’re requesting slash 28 by default. Again, this also could be a parameter that you pass in.

Karan Munalingal • 25:47

Hey, Johnson, as you were picking through, you know, a lot of our customers are already using it. Like, one of the initial questions that they had is like, Hey, it looks like I’m writing the English, but I’m also picking what I’m providing to the agent to go do the job that I’m asking it to do. It slightly differs from some of their other experience where, you know, Their current stack has access to all their MCP servers, where it essentially picks the right tool based on its own selection process. So, can you talk a little bit about why this was a chosen methodology within Flow AI and how it potentially ties to the decision making that the human has to put in the effort to say these are the only tools agency have access to? Can you talk a little bit about that?

Joksan Flores • 26:42

That’s a great point, Karan. Thanks for bringing that up. Yeah, we have made a very conscious decision of picking tools at build time, right? And we will be improving this along the way on the way that we recommend toolings actively to the builder. Now, this is very much a build activity today, right? We are very focused. We are attentionally very focused on providing self-service experience to our customers, whether that is they are requesting services via an API, via form, via et cetera.

Joksan Flores • 27:07

But also, because we live in the infrastructure, networking world, firewall world, and so forth, people are very, very, very risk averse, right? So, I want to make sure that anything that I provide to my agent in terms of tools, it has access to, right? You know, one of the things we always talk about, the contrast is, you know, we just like to poke fun, of course, nothing against open cloud services and things like that, but we all like to poke fun on those services because sometimes you give it 70 tools or 70 MCPs with hundreds and thousands of tools. And the lack of governance, right? It’s kind of scary that I might ask it to do something and it might just go off the rails and start creating it. It might decide that it needs to do something else in order to accomplish the goal that I provided. Here, I know very purposefully, Karan, and hopefully this answers and helps and you can add color.

Joksan Flores • 27:57

But the idea here is that I’m very consciously going to provide the tools that I know that my agent needs and the tools that I want my agent to use. If I just wanted the agent to go through the mock, you know, mock service creation, I wouldn’t give it the EC2 ad, which I will do today. I would just give it the list and not the ad, right? Because I know the ad is a destructive operation, right? It’s create EC2 instances. So, I have that ability to build and decide how it gets built. And then the person that’s coming and requesting services has no control over that, right?

Joksan Flores • 28:26

They just have functions, they just request the service, and hopefully, the agent has enough tools to do what it needs to do, but it doesn’t have too many, right? That’s the idea.

Karan Munalingal • 28:34

Yeah. And usually, what customers have also asked is within the prompt, if you know, obviously you mentioned, Jackson, that when you’re building your agent project, now I can share it with somebody. Let’s say you share it with me, right? And I go in and I start modifying your prompts and say, Hey, this is very, a very good start. I also wanted to start creating tickets and gather this information and send me an email about the specific information. But that might not be the true intent of the agent, right? I just changed the intent of the agent because I wanted it personally, but it doesn’t comply with some of the requirements.

Karan Munalingal • 29:11

And I think this is where, contextually, when you’re setting a context for the agent, just because you changed a prompt doesn’t mean it can actually go do that in the network, right? This is where the explicit tool association becomes a bigger deal because I can change prompts all day, but If the agent doesn’t have access to doing those things, it will never do them.

Joksan Flores • 29:34

100%. Something that’s very, very, very important to us, right? This is like, I always say the contrast between an engineer that only gets to do, you know, maybe tier one activities and they can only have permissions to, you know, create access ports or troubleshoot access ports versus the ability to go onto devices and shut down BGP peerings and those kinds of things, right? For our audience that does networking, that can relate, right? Because typically what you would want that engineer is to go and escalate that problem to somebody else. We can make our agents, we can design our agents to that very same image where they don’t have access, right? They can only read and recommend, but they can’t just go ham and do a whole business creative things, right?

Joksan Flores • 30:10

Yep. Okay, so we added a couple things while we were chatting there through the two controls. I’m going to add the last one, which is create VPC. And I think I got everything that go to my list. So I need EC2 list, add, VPC add, as VPC add. I got the wrong one here. I got a lot of stuff, right?

Joksan Flores • 30:28

There’s a lot of stuff built, a lot of people building here. There we go. So we got those and list, and then we got the info blocks and view data. Like Karan also said, right, these can mutate, right? I can come and edit my agent at any time. So right now we’re just building from scratch. We can come and add, remove, delete, and do all the things as needed later.

Joksan Flores • 30:48

Now, the next thing is a profile and model. So this is another important piece that we kind of glance over a lot of times. We have this concept of model registry. I think a couple of weeks ago, a few weeks ago, we did a webinar on bring your own model. And we explore this topic quite a bit, but essentially, I have profiles that have been enabled by my admin here. Let’s pretend that I’m not an admin for a 2nd . And I’ve only been given access to these few, right?

Joksan Flores • 31:10

So, like, my provider Anthropic model here only has access to Cloud Sonnet, where this guy here has access to a lot of other stuff, right? Notice I don’t have Fable, right? Because my admin doesn’t very self-conscious. He doesn’t want me to use Fable 5 and waste a lot of tokens doing these activities. So, Sonnet’s good enough for this. So, this is what we’re going to do. So, we’re going to do Sonnet 6.

Joksan Flores • 31:30

And then, also, runtime permissions for who gets to run this agent, Karn. Another super important thing down here. So, I also get to say which groups. Now, this is very group-related. So, this will be like my compute team. In this case, I’m just going to do solutions engineers. They get to trigger this agent.

Joksan Flores • 31:48

That’s it, right? As a workflow or as an operations manager entry, which is what we’re doing at build time, I can do anything here, but as we expose it, these permissions will govern what the platform has access, what anybody has access to. And also, down here, you got some kind of tidbits, right? Tool, tool, um, tooltip that says people need to have API write and read in our operations manager in order to be able to expose these, right, from a trigger perspective, which we’ll explore in a 2nd . I think that about covers the whole journey of agent building. Karn, you agree with this?

Karan Munalingal • 32:18

Yes, I think that was fairly straightforward. It was interesting for you to use existing APIs and an already available human-in-the-loop task. I’m interesting to see what you’re going to do next.

Joksan Flores • 32:31

Yeah, awesome. And that’s yeah, good call out too. So human in the loop, API call that’s existing on my info block system, right? My systems already have API lists is using those are automation. I also have scripts. So these are Python scripts, like I said, photo three scripts that are already out laying around. They could be workflows.

Joksan Flores • 32:46

They could be Terraform. They could be Ansible. Not picking sites, not picking tools, whatever works for your team, whatever people are using, bring them on and use them today. Okay, so our agent is ready to go, I think. So now we go and we say run agent here. It asks us for these variables. Now we’re not going to run it from here, which is validating that everything looks good.

Joksan Flores • 33:05

And we’re going to go and expose it because we’re not going to be here forever building. So let’s go ahead and expose it. And we call that thing, what do we call it? No, not that guy. This one. AWS app info provisioning. So I’m going to call, I’m going to, I’m in my operations manager and I’m going to create an automation.

Joksan Flores • 33:19

And this is how we expose as self-service. And this interface looks going through, you know, a journey of transformation. So this might look slightly different in a couple months. But for now, this is the interface that we get. And then in our operations manager, we have our agent. And I’m going to go and search for that. And my agent shows up right here.

Joksan Flores • 33:37

Now, I have my agent tied into my entry in operations manager, which is a self-service portal. But I’m going to go and create triggers as well. So I’m going to go create a manual trigger 1st . And this is a form trigger. I’m going to select it as manual here. And you can see that all the parameters that were built in the agent will appear here automatically, right? They’re rendered here.

Joksan Flores • 34:00

So the builder gets to decide what are the inputs for that agent, and my operations manager is just going to enforce it. And the other thing that I’m going to do, Karan, just for, you know, kind of for not losing the habit of doing this, I’m also going to create an API for this guy. So I’m going to go and do AWS app infra provisioning.

Karan Munalingal • 34:19

And Johnson, I think this is very important. As we kind of discussed early on, your consumers don’t just come through one channel, right? Application teams don’t like, and this is what I’ve heard directly from them, they don’t like filling out ServiceNow request form just to get a record or EC2’s instance spun up for testing purposes and then view their ticket and then have a closure. They essentially just want to call an API or drop an event and have somebody do the work and then move on with their life. So it’s a very interesting paradigm that we have a lot of our customers that have consistently asked: once I build something, how do I expose it in multiple different ways? Because my consumers, some of them want to use it as part of their pipeline and some of them are okay filling out a form. And some of them already have an application that drops an event in the ether, right?

Karan Munalingal • 35:15

So I think it’s very important, especially on the exposure side, that you’re nimble into how somebody wants to consume. Because what we have seen is if you lock it down, hey, this is the only way the adoption kind of goes down.

Joksan Flores • 35:30

Yeah, and Karen, right, very important, right? Not to be fooled or like just, you know, when you look at the prompt that we built there, as simple as that looks and as just minimal tools as it has, because it does, right? It’s got like five tools, a couple scripts, and an API call into InfoBlocks. That’s true orchestration, right? We’re touching multiple systems. We’re doing a VPC, we’re allocating a networking info blocks, we’re going to create an instance. This is something that people might request as a self-service.

Joksan Flores • 35:55

And guess what? Application teamers are really good at deploying things on their own, right? They have a pipeline running that’s deploying the thing. Why would they stop in mid pipeline and go and fill out that ServiceNow ticket? No, it doesn’t make sense, right? Put the approvals inside of the platform. Somebody go ahead and can approve the provisioning piece itself, but at least the request itself happens in terms of what the organization needs, not just forcing people to go and open the darn ticket, right?

Joksan Flores • 36:21

Okay, so I’m going to go and launch this thing. So let’s just pretend that now I have switched my phase and now I’m an operator and I’m going to request things manually for now. So let’s do it webinar demo app and we’re going to use a template that I already selected here. My template is demo and ODB. This could be a dropdown. I’ve just chosen to kind of pick these out manually. I’m going to put the requester there, which is my email.

Joksan Flores • 36:44

And then I’m also going to go, we’re going to go and provision this in US West One, right? Because it’s our MTS region there. So I’m going to go ahead and click run now. And we’re going to be redirected into agent sessions. And now our agent has started reasoning through the steps. And this is our auditing window, right, Karen? So this is where you can see everything that’s going on with this agent.

Joksan Flores • 37:03

So I’m going to go and hide that. I’m going to make that full screen. You can see that our agent has the entire prompt. The prompt has been templatized and so forth, saying webinar demo app is the app we’ve requested. I’ll work through, and this is back the reasoning trace from that agent. I can expose it to our tokens and everything mentioned here. Validation results, right?

Joksan Flores • 37:22

The most important thing for me, right? A lot of business stakeholders, engineers, and people are very risk averse, right? They want to validate 1st . I can have my agent fail fast and say, if you provided a name that doesn’t fit my standard naming for the application, I’m not going to provision it. I’m going to refuse or I’m going to go kick it back so that you fill it up again. I get to decide that at build time, right? I get to put that on my terms.

Joksan Flores • 37:44

So here we went through validation. US West, valid region, US West in California. This is our reasoning. I didn’t tell the agent what to do here. Demo no DB, non-empty, valid template string. So we have a template that exists in my service provisioning script. The application is valid for both the VPC and the EC2.

Joksan Flores • 38:01

And the requester, it’s all good, identified. No corrections needed. Everything is good. Now we move into a human in the loop. So this will redirect me to Work Center. Typically, the operator persona will just live in Work Center, right? In our case, we’re kind of playing two roles.

Joksan Flores • 38:14

So, we’re going to go ahead and split that up into two windows. I’m a big fan of this. And I’m going to go ahead and open my Works Interview. And you can see that I get here all the information, right? We were before looking at the agent trace. So, this will be safe for posterity, right? For logging and insights and all those kinds of things.

Joksan Flores • 38:34

A person approving, this is what they’ll see, right? So, imagine if the application gets requested via somebody else, they’re not going to see that. My approver gets to see this, and this gets to be assigned, and so on and so forth. And they get to see all the data and they say, Okay, this looks good. I’m going to approve and proceed with the deployment. All the validation results have passed, there’s nothing going on. And even the agent went an extra mile and said, Approving will trigger info blocks have node allocation, AWS VPC creation, AWS instance launch using template node, demo, node DV.

Joksan Flores • 39:04

Karan, you’ve done this for a while. Um, I didn’t templatize this, of course, but if we were doing this with a script or something, we’d probably have to go and create some sort of ginger template or something. And then now I have, it’s another piece of maintain, right? So, kind of cool that agents will give us some of these kind of cool bits that are impressive every now and then, right?

Karan Munalingal • 39:22

Yep, I think that’s where the reasoning comes in. The question I was going to ask, Jackson, is: this is great. Like, this is a 1st step where I feel I still have the control if our agent goes and does something in my own network. I think people have. Strive to keep that control from an expert standpoint. And over time, even in automation, they’re like, yep, I’m looking at the same data again and again. I’m approving the same thing again and again.

Karan Munalingal • 39:48

Move on, right? But I think when we’re coming back to agent, and you saw that the agent actually inferred some of the data that came back from a response and it made its own suggestion, it created its own template. Like it is consistently doing things that we had to type up in a Python script or a workflow, things like data transformation, right? Like I saw when you were doing your prompting, nowhere you had mentioned anything about HTML templates to use or a Jinja template. Like you basically said, visualize this for an approver in a sense that they can make a decision, right? Think about me not having to do that every single time I have a brand new service, or if I change my mind to say, not only am I going to have one human in the loop task, I’m going to have three after every step. It starts to add up on how much you have to manage and maintain.

Karan Munalingal • 40:47

But now, with reasoning, it is doing that for you. It’s not taking the action in the infrastructure, it’s just doing all the other non-functional tactical things like transforming data from one API to the next and validating certain things and presenting the view. I think this is fantastic, especially for folks. You know, who are looking at AI as an opportunity to leapfrog, you know, someone who is in doing little automation. I think we believe that there’s an opportunity for them to start, you know, improvising agents so they can take advantage of their existing assets, like APIs and scripts, and stitch it all together. This is great.

Joksan Flores • 41:30

Yeah, that’s huge, right? And not to take anything away from the people that want to do that, right? They can do it. If you want to templatize this, right? I didn’t tell it to use. We’re going to see the next one, it should be a little bit prettier because I gave it guidelines and I said go and build it in an ISO way and so forth. But you could go ahead and template it if you wanted to.

Joksan Flores • 41:45

We provide all the capability to do it. But just to me, getting started super simple with the assets I have, Karan, like you said, it saves me a heck of a lot of time. So, right, that time to value is super important. We’ve done this before, right? We’ve done it for a long time. We can create workflows that basically replicate the same behavior. It takes me a while longer to build a workflow, right?

Joksan Flores • 42:01

I can just get time to value here very quickly. And then, as I go along, I can make things more agile, I can change them to my test or whatever, right? But this lets me go way quicker and start delivering services to the organization. So, I’m going to go ahead and approve that. And you will see the transition on the left here in a 2nd . And we’ll let the agent do its thing. I was going to go reasoning through, right?

Joksan Flores • 42:22

Approval granted, moving to step three. Okay. So now it went ahead and failed. And it reasoned. This is another thing, right? This is also, a lot of times we have mechanisms to address this so that it doesn’t fail ever. But it’s kind of nice to see that it failed and it realized that the container block needs to be referenced without the prefix length citation in that field.

Joksan Flores • 42:42

I’m going to go ahead and retry the tool call and do it again and do it right. So it did it. And it’s moving as it goes along. So it always, you know, kind of cracks me up. And people appreciate that. It just recovers automatically. It’s doing its thing.

Joksan Flores • 42:54

It’s just doing it.

Karan Munalingal • 42:55

And Jackson, as this is going, you know, one of the questions that I’ve gotten from, you know, customers in general is: hey, it seems to kind of think through and take the action on retrying. And if it sees an error, either it’s going to pause. How much control do I have as a developer of this agent? To actually say that, hey, only try it two times. And the 3rd time it fails, I want you to exit and create an incident ticket. Like, how much control do I have as a developer? So it doesn’t seem like it’s taking actions without my permission or intent.

Joksan Flores • 43:33

I can control all that through prompting, Karan. And then we also have controls on the platform of how much it can do. But I can tell it, right? Sometimes, like, we’ll go through. I’ve gone through this during testing where I’ve said, sometimes. So, yeah, Karan, I have lots of controls from the tooling retry perspective. I can tell it, hey, go ahead and retry tools one, two, three times max.

Joksan Flores • 43:54

We also have several controls, but typically, you know, one of the things that we do is we want to avoid some of that stuff, right? So, whenever some of these happen, what I would go and is I would go and we can explore that, right? Or either a different time or this time. If I’m passing it, if I’m giving it values that are not important and not this one, this is a bad example, right? Because that was just a timing thing. But this one, e.g. , here, I have the ability of creating decorators on our platform, which is essentially telling, letting the agent know to. Teach the agent, hey, this is what this tool is expecting from an input perspective in an exact way, right?

Joksan Flores • 44:29

Like essentially, the error when it happens here is that this slash shouldn’t be here, right? It should be 28 as a network size. And sometimes the API spec or the tool as it comes from the vendor, it’s not super explicit on that. But I can go ahead and decorate, right? I can go ahead and overwrite the API guidance so that it’s very specific and tells the agent, hey, don’t use a slash, just only pass in an integer, which is what’s required in that string format. So we have a lot of those value-add features, but we can definitely have a lot of control in the deep down on which pieces to do and which pieces to not do and how to retry and so forth. If you look at here on the right, though, it got to doing my little template that I had told it to build.

Joksan Flores • 45:13

So it’s got the deployment has been successful. All six steps have been completed. And look at that, Karn, the level of detail, right? Like I told it, hey, go and capture some of the things that were provisioned along the way, right? The InfoBlox container. Obviously, these are input parameters. But then if we go through the approval log, pre-deployment approval, approved, this is the timestamp.

Joksan Flores • 45:33

So it got approved that this time UTC, it’s got the subnet allocation specifically for this use, right? 1099.1.80. I got the usable host, the network reference from InfoBlox has captured the VPC parameters, right? The VPC ID subnet ID. So I could do all sorts of things after the fact with this information, right? I can go ahead and create a record of this provisioning, which is, you know, another webinar, right, for a different day. But we could go and create records, and now I can maintain crowd activity operations and things like that.

Joksan Flores • 46:05

So a lot of detail here. And I just told the agent, Karan, like you said, go and create me an HTML report. Capture the I told it, I capture the IDs of everything that gets provisioned and then report it back as a human in the loop activity. What do you think of that, Karan?

Karan Munalingal • 46:20

I think this is phenomenal. You can think about. You know, when you’re building, you know, one of the trends that we have seen for the last 12 years, Jockson, is when you’re building an automation or an orchestration, before you go build it, you actually had to go write down every step that you wanted the automation to actually cover. On the contrary, now, when you’re looking at something like Flow AI that has the ability to integrate into LLM for reasoning purposes, you essentially have the ability to just lay out your intent without specifying everything today, right? The idea that I just got is like, hey, if it’s capturing all this information, can I now go store it somewhere else? Like, can I just go back to my original prompt and say, hey, all the details that you have gathered, organize it properly and then store it in a space that I can reference in the future, right? This is, it gives, like, based on working with customers during the innovation program, I think it has given them so much assurance that now they can change their mind every single day without having to panic to say, oh man, what’s going to happen to my existing workflow or script?

Karan Munalingal • 47:34

How do I have to reconform and then make sure that it doesn’t break anymore? I think this now gives you an opportunity to trial, trial, trial until you’re happy very quickly, until you take that all the way into production, right? That’s what we’re talking about. How do we take someone’s intent and make sure that the intent is perfected as you’re taking that into production very quickly? And this is great.

Joksan Flores • 48:01

Yeah, I’m going to go ahead and acknowledge that. And Karan, to your point, right, and we’re going to go finish that session off. And while the agent reasons through and thus conclusion and all that stuff, right? It’s that’s 100% true. And a lot of these intents are already expressed, right, Karan? Like, this is not new. People have been doing this for a while, right?

Joksan Flores • 48:16

People build MOPs. They might not build agents, but they build MOPs. They build procedures. They build confluence documents. I was talking to somebody today, one of our peers today, and said, you know, I have been through a lot of finance customers and customers that deal with securities and all those kinds of things. And there are CSA organizations putting policies on how provisioning needs to do for needs to be done for certain things. A lot of times they’re just PDF documents that are issued or they’re a confluence page, but these could become actionable assets that you can use for building.

Joksan Flores • 48:48

You can take those things and convert them into part of your prompt, or you can make them into a skill or what have you, another webinar for another day. But you can make those into your prompts and now ensure that your execution is following that policy that you’ve been kind of told that this is what you have to follow instead of it being subject to interpretation. So this is definitely a big aid for a lot of people. And you still have the guardrails, right? You still, as an engineer or as an approver, you’re still sitting here watching and monitoring everything that’s going on.

Karan Munalingal • 49:18

And Joksan, I think my next question naturally is, hey, I just built it. Is it perfect? Right. We have to tell the customers like, no 1st prompt is ever going to be perfect. Right. But if no prompt is ever going to be, no 1st prompt is ever going to be perfect. The 1st agent you’re going to load is not going to be perfect.

Karan Munalingal • 49:39

Then tell me how easy it is for me to go modify so I can make it perfect. Right. So it would be awesome to actually see. you know, you changing your intent because you learned something new or when you ran this into production, the production environment is fairly dynamic, right? Things are happening independent of you wanting it to happen. So when that happens, how do you now address that rather quickly to say, hey, I never anticipated seeing this issue. Can you go address it now?

Karan Munalingal • 50:08

And I’ve seen, you know, me and you have worked in the field for a while. Imagine that happening. That becomes a requirement. Now you have to go modify your existing workflows or scripts to address that specific scenario. How easy it is for me to go modify the agent that I just built?

Joksan Flores • 50:26

Yeah, and in here we tried to make it super easy, right? Last week, I guess, like I said, a couple of weeks ago or a few weeks ago, we were talking about how to switch models. That’s a very easy change. If I wanted to switch the model that I’m executing, I just come down here, change my model provider, right? In this case, if I wanted to go from Anthropic to some Olama or open router run thing, I got all this stuff, right? Metamuser Glimmer, it’s all the rave. I could go and pick models, but I’m not going to make that change now.

Joksan Flores • 50:52

I’m actually going to go and modify the prompt. I’m going to do make a change in our behavior, in our agent behavior, and we’re going to reorder some steps, Karan, because we, like you said, we tried to make it super easy so that people can modify. And it’s like, hey, saying now, instead of you going through, you know, perhaps the approval still holds, but we’re going to go and add another step here to say, we’re going to go ahead and let’s do this. I’m actually going to go back and let’s add some sub-step and I’m going to call it 2A. And I’m going to go ahead and say very creative here: create a ServiceNow . And I’m typing this one because I didn’t. This is all here on the fly now.

Joksan Flores • 51:32

So create a ServiceNow , change request. And let’s do that in, let’s say, add create, create, cr record, and add all the parameters for the request and approver information. So we’re going to go and do that, Karan. And then what I’m going to do is also, I’m going to add another step at the bottom down here. So let’s go. I did 2A right here. And I’m going to have, you know, let’s tab it up a little bit.

Joksan Flores • 52:05

Make sure that it looks okay. I’m not sure if it’s going to look good. My markdown is not great, but let’s go do that. And then at the bottom, when we confirm deployment, let’s see, launches into instance, confirm deployment. Once we confirm deployment here, let’s do 6A and we’re going to do update. Update. Update servers now request because my policy now changed.

Joksan Flores • 52:25

And I just want to do, instead of just doing it all locally in the platform here, let’s go update. Update CR. Update CR record with all. With all provisioning results and assets. Sure, I think that’s good. Okay, so we’re going to go and update, create a CR and update it. And now I got to provide it the proper tools to do that.

Joksan Flores • 52:51

So that’s change request. We have a tool from ServiceNow that is create change request. And we’re going to go ahead and actually decorate it, Karan. And we already have this decorator here, which essentially is like I explained earlier, right? It overwrites the behavior of the API. ServiceNow’s API is a little bit generic. So in this case, we’re just going to go ahead and use those decorators that have been created here before.

Joksan Flores • 53:16

And the same thing for update change request down here. And I’m going to decorate this one as well.

Karan Munalingal • 53:22

And Johnson, these are native APIs that you’re associating provided by the vendor, right?

Joksan Flores • 53:27

Correct. Yeah. Another, yeah. Sorry, you’re bypassing that. Great point. These are native provided by the vendor, like Karan stated. They come from the API spec, loaded into the platform, and bada bin, bada boom, right?

Joksan Flores • 53:39

As it came from the vendor, that’s how they show up here. And that’s how I’m going to consume them. So let’s just go ahead and do that. I think let’s look at that prompt, make sure it looks okay. The spacing looks a little off, but hopefully.

Karan Munalingal • 53:54

So you essentially, in about two sentences, have incorporated change management to your original agent design, bro. So now the agent not only talks to Infoblox and gets the IP, provisions the EC2 instances, leveraging your existing scripts, but now it’s incorporating change management so it can audit what it’s doing.

Joksan Flores • 54:17

Yeah, 100%. And, you know, very quickly, right, the tooling was already available in there and everything as long as it’s integrated into the platform. And my paradigm still remains the same, right? I’m going to improve inside of the platform using our work center, even in the loop activities, but now I’m going to record it all. So let’s go and run this. Just go and refresh, make sure.

Karan Munalingal • 54:40

And what’s interesting here is from a consumer standpoint, because we’re talking about taking agents to production, right? The last thing you want is just because you wanted to add more capabilities at the agent level, it started impacting your consumer. I see that my exposure, the inputs that’s required of me as a consumer remains the same. But now, on the back end, when this is my request is launched, you’re basically going to capture that in ServiceNow in addition to what you were doing before.

Joksan Flores • 55:11

100%. Yeah, let’s go back and see what it does. Yeah, nothing changes in the front end, like you stated. Let’s go back and let it do its thing. Okay, same activity again. We’re going to split view, go to work center. I collapse these two, go by preferred view.

Joksan Flores • 55:35

See, now we’re going to go back and allocate, blah, blah, blah, do its thing, create VPC, launch instance, everything else is the same. My name is different now so that I don’t clash with anything that I had created before because the instance is already there. NAWS is not as important.

Karan Munalingal • 55:55

Oh, that’s interesting.

Joksan Flores • 55:56

My creature happened very quickly, immediately there, Karan.

Karan Munalingal • 56:00

And it formulated the payload by itself, which I struggle doing so in my own script. Every time a vendor changes their API, I have to go touch that script again, you know.

Joksan Flores • 56:11

And it’s doing its thing in the background there. And actually, I want to go and look at not that I don’t believe it, but you know, change, work, change. Oh, I noticed I didn’t tell it, right? I could go, you know, look at the tickets right here. So I could go and actually give it guidance about how to name the short description, which is the one that shows on the CI table here for change. And I didn’t. I just, you know, we just did it very quickly in 2 s when you prompted me.

Joksan Flores • 56:39

And look, AWS VPC plus EC2 provisioning, webinar demo, app WCR in US West One. And if I go back and open that ticket, let’s see, I have to double-click this thing. See, all slow. Okay, here we go. So, we got some provisioning details requested. Oh, look at that. Perfect.

Joksan Flores • 57:00

So, look at that, Karen. In the description itself, we got self-documented. And again, granted, I can add guidance to this to what things to capture. But it’s doing, I mean, it did enough. I think this is good for me. Validation results pass, no connection required, plan steps, allocate, blah, blah, blah. Do its thing and do everything else.

Joksan Flores • 57:19

And then we’ll see eventually, we’ll see notes and things like that.

Karan Munalingal • 57:22

Oh, look, it just came through. It’s very impressive that, you know, I don’t have to now coach somebody to humanly put that in here, right? It only took two sentences and two associations, and it started doing more than what I do today. Yeah, thinking, you know, this is great.

Joksan Flores • 57:40

Yeah, take an enrichment alone, Karan. You know this, right? You talk to a lot of people. A lot of times typing this stuff becomes exhausting. People won’t do it. The machine doesn’t get tired. That’s why I like to tell my customers: the machine doesn’t get tired.

Joksan Flores • 57:52

It gets paid for this. So, you know, having the richness and the evidence there, if it serves now as your SOT for change, then now you can do this. And bring your tool, Karan. You saw me at the API. We could as easily tomorrow, the organization wants to move to FreshWorks, swap a couple APA calls, and off we go.

Karan Munalingal • 58:12

And we have a lot of customers, as you know, Jockson, who have spent a lot of time in writing some of the playbooks, whether they’re modular or not. I think there is a great opportunity to get started with improvising agentic capability within infrastructure without undermining that the agent is going to make changes, right? It’s still your deterministic tools that our platform actually leverages to make those changes, which is a critical part. It’s the same API, the same playbook, the same script that folks have been using and feel comfortable with. An agent within Flow AI does the same. It uses the same methods to actually make changes. There’s nothing different there, which is a big part, right?

Karan Munalingal • 58:57

You have already built a governance, you’ve already built the security mode around it. All we’re saying is use what you have so you can go faster.

Joksan Flores • 59:04

Yeah. And this is the same, right? We just took a bunch of things that are laying around and we just did augment it, used it, proved it, augmented my script, my mode of provisioning hasn’t changed, right? Let’s assume that I’ve been doing this for years from my laptop or from a runner server somewhere. Now look at Carn, like we just added the CR thing and in summary, it just has ServiceNow CR. And this could become an email, right? This could become an email with the change ID, the approvals, and all those kinds of things.

Joksan Flores • 59:30

Everything is in here.

Karan Munalingal • 59:32

This is great from an auditing standpoint. Imagine being able to keep these sort of records so you can always associate which application was instantiated for this individual developer. What were the associated IP addresses? Who approved it? I think all of that boils down to. Keep in a hygienic way and a standardized way of capturing this for any infrastructure change. It doesn’t have to be app info provisioning.

Karan Munalingal • 59:59

I can now take your prompts, use 70% of that, and just insert the other two pieces of my F5 changes that I’m making, but everything else remains the same. You know, that’s fantastic. I think that was a great demonstration. Thank you, Joksan. I think you gave us kind of like an ability to immediately leverage an agent harness and your existing tools and start building very rapidly. So, you know, one of the things that We always talk through is a lot of our customers and folks in the industry have been on this journey where, you know, they’re looking at, hey, do I get started with automation 1st and then go to orchestration?

Karan Munalingal • 01:00:46

I think we have an opportunity with Flow AI. A lot of our customers have the ability to start with agents 1st , right? Bring your SMEs to the table, you know, enable them with their existing toolings and their words, their mops. They can quickly turn out these agents that are doing real things within your infrastructure with a lot of guardrails, governance, security, as well as humans are always in the loop as you’re starting out, right? So when you’re starting with agents, don’t feel like you’re exposing your infrastructure to agents running wild. You’re not. This is a very governed platform where you can go fast, but over time, as you’re building more agents, you start improvising more validation, compliance, standards.

Karan Munalingal • 01:01:33

And ultimately, you have the ability to go full-on orchestration. What Jackson did not get to today is now the agents that you’re building that are doing specific things, you can now incorporate that into your deterministic workflow orchestration to drive more value, right? Because now you’re combining a combination of thinking, which is reasoning, along with your very compliant. Steps that you must take to change infrastructure within your environment. Right? If you look at the bottom, Human in the loop, still very important.

Karan Munalingal • 01:02:06

Human on the loop will remain important. And as you kind of progress with more assurance, now we’re talking about multi-agent systems where you have a combination of agents that you’re building with your existing deterministic capability and some new stuff that you’re going to build rather fast and combine that together to get accelerated value, right? So, with that, you know, I would like to thank you very much, Jackson, for doing this webinar with me. That was a fantastic demonstration of the capability that will drive time to value for our customers, as well as the ability to bring more people, engage more people with their expertise to start taking an agentic path forward and then lean into the orchestration longer term. So, you know, some parting words for you, Jackson. Anything that you would like to share for our audience?

Joksan Flores • 01:02:59

Yeah, I mean, if I were to take a leap here, I think this is something that would be very comfortable, assuming the platform is deployed and the guardrails or the integrations are put in. I feel very comfortable deploying something like this in something like a week, right? Doing some testing and deploying. That’s the idea, right? If you go back to the title of the slide, from prompt to production, let’s get you there quicker, faster, all the guardrails you want. Maybe you don’t do the provisioning car. Like you said, maybe you start small, read-only, but you can get to production quickly.

Joksan Flores • 01:03:24

You saw it here. We just did changes on the fly, added functionality. And, you know, this is just great technology. You got to take advantage of it.

Karan Munalingal • 01:03:32

Awesome. Thank you, Joxin. And thank you, everyone, for attending. Looking forward to more of these webinars with you, Joxin. And hopefully, everyone enjoyed the content. Bye, buddy.