How to Host Agentic Workflows in the Cloud (3 BEST WAYS)

Hey, in this video I'm going to show you how you guys could take your Agentic workflows and put them in the cloud. Essentially allow you to call them via web hook, schedule them on a trigger so they can run without you. Um, call NADN and Make. com web hooks through your agents. Do the inverse. So call your agents using make. com or NAD HTTP requests. I'll even show you how to set up local servers. Everything I'm going to talk about here is going to be fast. It's going to be really cheap and most importantly pretty reliable as well. So the first way to do

this is to set up web hooks or what's called cron which is basically just auler in a service called modal. Now this takes just one click and one prompt for the user. It's very simple and I'm actually going to give you guys a big comprehensive document you can feed directly into your agent to do this for you. It runs on a serverless backend which is very cheap. That just means you don't have any ongoing costs. It's just a build almost like per transaction. The server spins down when it's not in use. You get a web hook URL. Then you can also stream what's called chain of thought to Slack or whatever for logging. I'm going to show you all this in a moment. Modal is this service over here. I'm not affiliated whatsoever, but there are a lot of big companies that use this. And the real benefit to this is they just do what's called cold starts really quickly. Um, essentially, uh, a lot of the time when you run a similar sort of structure when you send a request to a node that hasn't been accessed in a really long time, takes quite a while to warm up. And so this might take 20, 30 seconds, hell, maybe even a couple minutes for a simple API request. Um, if for our agentic workflows, I'm sure you can imagine timeouts are really important because agent logic takes a little bit longer to go through than typical, I don't know, scripts or whatever. So, we want to minimize that as much as possible. Modal does that for you. I mean, like I get to call things that are inactive and then I receive or returned a request in like a couple of seconds. So, it's very powerful. The way that it works essentially is you get to set up web hooks directly in the cloud. And every one of these web hooks is basically just a fully functional agent that has access to whatever tools you give it. And so you can see here I have the create proposal from transcript web hook. I have the directive web hook, general agent web hook. You know, I got a bunch of these. Um, it also allows you to set up schedulers. So you can see that with this little clock icon, but this hourly lead scraper is basically just scraping leads for me every hour completely autonomously and then it's dumping all their information into this big sheet. So lot that you could do with this and I'm going to show you guys everything in a moment. But suffice to say, um, not affiliated with these guys whatsoever. They give you $5 in free credits. I've probably sent I don't know like several hundred requests over the course of the last few days and I've literally used 1 cent. So this sort of thing scales really well and I just wanted everybody here to have access to it. Okay, so two main ways I'm going to talk about today. The first is you just running specific executions. Now when I say specific executions, I'm referring to my directive orchestration execution framework. In case you guys don't know, if you're running um agentic workflows, you essentially have some structure like this. You're going to have highle instructions that tell the agent what you want it to do. And in my case, I always store it in this directives folder. Um, these pieces of information are typically pretty straightforward. They're just like, hey, this file creates a panadoc proposal for a client. Here's how you do it. You start by gathering information. Then you generate the content. Then you actually go through um a Python script that we've developed and then do the execution. And so when you separate concerns like this, like highle instructions from the actual like Python script and sort of like computer logic going on under the scenes in this executions folder, the model is really flexible, but then it's also pretty reliable because it just ends up doing the same thing over and over again. Contrast this with always giving the model different instructions and so on and so forth. It's a mess, right? Not usable in any serious business application without this. So if you use option one with specific executions, basically what you're doing is you use the directives to build the modal web hook and it's just a one-time thing. And then the modal web hook is just executions. So this is very similar to just setting up your own API. But because you know you've already written the directive and stuff like that, the agent is capable of writing these APIs super quickly. And so basically everybody on planet Earth is now an API developer uh when you combine this with the little document I'm about to send you. And the benefit to this is it's really fast. I mean these things run basically immediately. It's also super reliable because it's just basically Python scripts and um you know it's also super cheap because you're not consuming additional tokens or whatever. Now I'm going to contrast that with the second approach in a minute, but let me actually just show you guys an example of this working. So I have a little um NAD workspace set up here. I want you guys to know you can call your web hooks however the heck you want, but obviously the whole point of this, the whole name of the game is we want to be able to trigger these automatically, right? So we need some service to basically call it. And what I have here is I have this URL. Okay? And the URL is, you know, Nick whatever claude orchestrator scrape leads. mmoal. run. And then I had this question mark query. And this query is, you know, dentists and location equals United States and limit 100. And so what I've done here is I basically set up my own API using these executions with this agent. I'm just going to click execute step and it's going to go and it's actually going to run this in the cloud and return me some data. Now, as I mentioned, cold starts here uh mean that you know, typically there's a very long period of time before you call any sort of service where your server is idle. But modal occurred super quickly. I did that in real time, so hopefully you guys could see that was just a few seconds. You know, edit end returns you this kind of sexy looking JSON that's something like this. You have um you know, status, message, sheet URL. This is where all the leads are stored for my specific workflow, which I'll run through in a second. The name of it, and then really interestingly, you have the workflow itself, which is the agent basically breaking down the steps that are going to be taken in this execution. And so, you are storing, you know, the equivalent of Python scripts here, but there's a little bit more going on under the hood, which is pretty cool. So, if I just paste this sheet in, you know, you can see this thing has already gone through and done my lead scraping. Now, you know, in my case, I was only scraping 100 leads, but um there's more going on under the hood as well. It's not just doing lead scraping. It's also enriching. So, that's what that little update was a moment ago. It's finding some additional emails. It's personalizing and so on and so forth. And so, in this way, you basically have, you know, all of the information involved in the um highle instructions just used to build this web hook one time and then other services and stuff like that can just call this at whim. super cheap, super reliable, and probably the recommended approach for most people. That said, the internet is getting really cool, and so I wanted to talk about um in this case, just a proof of concept, but something that I think will quickly become the norm when it comes to quering these agents. Now, this is natural language as opposed to you doing some sort of really specific formula with limit and industry name and stuff like that. So, what we do here is we have our web hook URL and then we just have a question mark query. This is a query parameter and then we just give it instructions and we say, "Hey, run the proposal generator on my last call. " And the idea is because we have an agent that is living on the cloud, it is just awaiting instructions. When you query it with, hey, run the proposal generator on my last call, it gets the natural language and then it interprets which directive you want to call. It calls specific directive and then it actually runs the thing you want with its uh full power of flexibility. And so if I give you guys an example by calling in NAN, if I just go down here to general cloud, so you guys could see this URL that I'm calling is now uh same sort of thing, but the query is send an email to Nickleclick. ai congratulating him. I mean, obviously this is pretty vague, right? like there's no uh Python script or whatever that'd be able to do this thing in real time. Not unless you parsed out literally every word. But um anyway, in this case, this took a few seconds and it returned success. The query was this. And the response is email sent successfully. I've sent a congratulatory email to Nick with the message ID, whatever. So, I can actually go to my mailbox and I can see the message came in just a few seconds ago. It says, "Hi, Nick. Congratulations on your achievement. Best regards. " Now, I could have given it way more context and detail. I could have told it to do something else. Obviously, it would have. And this is just a really simple proof of concept because all I'm using is an email tool here. But hopefully you guys can see that adding uh flexibility in the ability to communicate with these things will soon allow us to build like this network of agents that talk to each other in pretty much just natural language asking things to happen. And in this way, you can build like much more complicated workflows where it's like sequences of agents talking to each other. Hey, I just scraped the leads. The link is here. Can you do this? Then the lead enrichment agent is like, "Yeah, for sure. I'll do the enrichment. " But I'm also thinking we should start reaching out to them. So, hey, here's this data. um could you reach out to these leads? You don't have to like code everything anymore. be really precise with syntax. The second way, and I really

like this because it helps people that have no code backgrounds integrate with this Gentic Workflow stuff, is to build hybrid directives. Basically allowing you to integrate like NAND, Make. com, Gum Loop, Lindy Flows directly into um some sort of IDE. So, it's cool because you can really easily integrate it with any pre-existing builds. Like a lot of people's intellectual capital has been caught up in like these traditional builds that they've built over the course of the last couple years. No use in wasting that or laboriously rebuilding that in some sort of agent platform. You can just use it. Super easy. All you do is you store info about the web hooks in your directive. Great middle ground. And the downside is, you know, at least the way that I'm going to show you today. It's not runnable on a schedule since you're calling from the IDE itself, which is right over here. But it is really straightforward to set up uh automated workflows like I showed you earlier where you have a web hook coming into an agent. Maybe the agent is what calls the thing, right? So, still triggerable, just um not in the way that I'm going to show you guys today. Okay. So, what do I mean by this? Um let's say you have a bunch of intellectual capital stored up in a workflow here. It's a Google SER/MPAP scraper. This specific flow, you know, was just connected here a moment ago and I just disconnected it. Basically just um scrapes a bunch of details for I think it's like Calgary plumbers or something cuz I live out in Calgary. And then it translates it to Markdown and then it uses AI to extract info and it dumps it all into a big Google sheet. And let me find the actual Google sheet because it is massive. It is a really, really big boy. This is more or less what it looks like. I guess I lied. It is not massive. Um, but basically you end up with a couple of records. page titles, page metadescriptions, like you know, it's like a scraping build, right? And obviously, you know, I want to do something with this. So, um, how do I take this and then make it, you know, accessible? Well, all I do is instead of connecting via, you know, some other sort of trigger like a schedule, whatever, I just use a web hook. And so, I connect this web hook directly to my flow. Okay, I copy the URL. I go back over here and u let's just set up uh let's set up a new directive. Call it new file. And I'll say call cool Google scraper. m MD. Okay. And now what I'm going to say is um hey, let me just use voice here so it's not really silly. Hello. Thank you. Um hey, I have a workflow set up here. I want you to call this URL with a get request. And I'm going to paste it in. Then I'm going to say this is just for demo purposes. You don't need to include any query parameters or anything, but hopefully you guys could see how you absolutely could if you wanted to. And now I'm just going to go and make a new cloud code instance here. And I'm going to say, hey, call uh let's see, cool Google scraper. md. And uh let me now actually run this. Let me show you how it works. What we're doing now is we're just weaving in the you know, N8 flows just directly into this. We're making them accessible. uh if you guys want to really quickly and easily create any sort of like documentation or directives for these things like highle instructions on how to do so it's really easy you literally just copy this whole thing okay you can actually copy all of the JSON here and you can say great here is a bunch of uh JSON that represents the NADN web hook flow that I set up that I just asked you to call I'd like you to create a highle directive my goal is I want to use you the agent to orchestrate the calling of this web hook I just want to make sure that you do it right so Create a new directive. Call it google_p_scraper. md. And then uh include all of the information that you feel is relevant based off the framework. You guys didn't know I'm just using a voice transcription tool. Many of them down here, but that's what this little bubble is at the bottom of my screen. Obviously, when models get really smart, you don't have to be anywhere near as concise or precise with your prompting. So, I could just talk at the speed of my thought. And it does a pretty good job. Look at it. Even told me nice workflow. Now, it's asking me what do I want to do? And I'm just going to say create a new directive that describes how to call this context about it etc. And if we go and check back in on our flow, you can see that what this did is it immediately returned that 200, but the flow itself is actually still occurring, right? So markdown HTTP request, it's actually running the actor. Now what we're doing is we're actually feeding things into the AI. It's kind of neat because we're actually feeding the whole website in. But um anyway, I'll tell you more about that later. Once it's done, you can see it just added some items to the sheet. I think it was this one over here. um we now have you know an additional record and contains a bunch of information about the companies and so on and so forth. What I mean here is we just trigger this um using an agentic workflow, right? Like we didn't actually need to do anything in NA or make and in this way we're basically just capable of triggering whatever we want. Now more on this um NAN recently unveiled like an MCP feature model context protocol. So you don't actually even have to like pass the web hook URLs directly. You could just use NAND's MCP. I haven't spent more than maybe 10 minutes looking at it. So I'm not going to include that feature here. But if you guys want more on that then um yeah definitely just use an MCP. Orchestration isn't the exact same. Uh, I find that sometimes, you know, you're loading way too much into context, but it works fine regardless. If we check back in over here, you can see we've now created this big long directive, which just includes tons of information all about what the workflow is, how to call it, and so on and so forth. And in this way, we've now blended a no code tool with our agentic workflow builder. Last section I'd like

to talk about is you can also run all of these flows locally in your computer and you can expose via Enro. Um, I should also say, I don't know, Cloudflare or whatever other service you want. In case you don't know what that means, that's okay. What we do is we set up a server on our computer which is just always running. Whereas with modal, you know, it's like instantiated when we need it. In this case, you just have it on your computer and then you basically allow your computer to be called by other computers um or other services or automation platforms or whatever. So, it seems really scary, but it's actually really simple to set up because again, you can just like feed in a big long prompt just like we did with the modal cloud execution framework and then the model just can do it right the first time. This is neat because you can set up cron which is scheduling and also web hook functionality. It does require a computer to be on 247. So if you're, you know, somebody that's doing this professionally and maybe implementing it on other people's businesses or whatever, you know, they would need a server and this starts getting into like deployment land and territory. But I think just for like demo purposes and getting people up and running and used to servers and building um cloud agentic workflows, it's pretty reasonable. So that's what I'm going to talk about next. So all I did in order to set this up myself is I said hey I want to set up a server locally on my computer to call a specific directive in my case upwork_crape_apply. md and I want to be able to you know do this from external requests. So set this up and then expose the URL for me and what I ended up getting is this URL down over here which if I open is ear-drove-passing. trycloudflare. com/webhook/upwork-scrape. Maybe this seems really intimidating to some of you, but this is literally just a URL that I can call or I can um you know, send a request to at any time that will run my workflow. And it's like, you know, technically it's locally, but I'm still accessing it through the cloud. And so you can expose this to other services to do things on your computer really, really quickly. And so what this is this is an Upwork scraper, right? Um so in my case, what I'm going to do is just click execute step. So it's now going to send a request over to this and then return the results. And less than 2 seconds later, this thing finished. As we see here, we have a sheet URL. The status was successful. Um, so what is this thing going to do? This is going to include my scraped Upwork jobs. So, just copying this over, pasting this into a new tab. We can see here just because I was running a test, I only did a couple of records, but we have what looks like some actual scrape jobs, virtual assistant for cleaning business, manga styled web comic writer, and story developer. Right, this is for like a writing keyword. It's very general and vague, but um, yeah, got the cover letter right over here. Got like Google Doc proposal, which is kind of neat. and I had this thing write like a supposedly in-depth personalized Google doc proposal. Hey, I spent 15 minutes putting this together for you. In short, it's how I'd structure your dual business VA system, so on and so forth. So, pretty cool stuff. Um, and I run this entirely locally on my computer, so it doesn't cost me a scent and I can just have this on. You do have to, you know, ask the question like, well, if you're running this on your computer and you're making it web accessible, like what is the real purpose? Why don't you just do it on your computer? The real purpose for this is just the web hook calls. So you can just set up another service to send something to your computer and then expose your like server link via the um in my case try cloudfire service but whatever you want. Um and I'm going to include like a link with a similar prompt just like the modal cloud execution framework here that you guys could use to set this up in seconds. Okay so I've teased this a couple of times but let's actually set up web hooks and cron inside of modal. Um this is the complicated bit so I'm going to walk you through this and then everything else you guys should be able to do just based off of what I've shown you in this video. I have referenced this cloud execution framework a couple times. What is this? This is a monster prompt that I've developed that allows you to just like feed this into AI, then have AI set up your modal for you. So, what I will do is I will simply copy this entire thing, okay, into a VS Code environment. Just going to go bypass permissions. Paste this whole thing in and I'll say, hey, I'd like to create a cloud modal instance for the let's do YouTube outliers. md directive. Follow the above steps. let me know if I need to do anything. Now, in my case, I've already set up all my like environment keys and stuff. You know, I've already actually signed up to modal and whatnot. So, um this is going to be a little bit faster for me than it probably will be for you. There will be some issue me asking it to do this the first time, but that's okay. The whole idea of the framework that I always talk about, which is having directives and then orchestrating the execution of various Python scripts, is this kind of self- analing, right? Like the agent kind of makes some errors and then it'll be like, "Oh, okay. Like, I made an error here. Let's fix it. " and then it'll go through that process over and over again. And I just wanted to show you what like the end to end process looks like. I know nothing about deployment to be clear. I did some front end way back in the day. You know, a little bit of backend to build um an app for one second copy forever ago. But like I'm not a good or competent developer by any means. All I did was I had it do this correctly after maybe 30 minutes of back and forth. Then I had it documented in that big cloud execution framework. Then I'm just giving all of that information to this. Okay. And sorry, I'm just realizing I already created it a moment ago pre-demo. So, let me just tell it to delete the old one and make a new one. That way, it's a totally fresh implementation. So, what it's going to do first is just delete said old YouTube outliers code. This is just a workflow that I set up a while ago that um gets a list of outliers in my niche for like specific uh you know, requests just so that I can keep an eye on you know who is absolutely crushing me on YouTube. Okay, so we did run into an issue. I told you I was only allowed eight web endpoints. We have nine. So, the current endpoints are all these. Which ones do I want to remove? Right? I'm uh keeping this in the video because I just want to show you guys like it is very much a very simple back and forth. It is asking me a very valid question here. Probability that anything is going to work on the first try is quite low. But I mean same thing when you work with like a junior dev or something, right? Like obviously you're going to have to go do some back and forth. So I don't know which one do I not like. Uh I don't like test email. So let's say remove test email. And I actually even recommended that. Okay. And I now have an endpoint available for me. So, I'm just going to copy this. Uh, let's see what is a good demo example. Yeah, let's just run this one here. So, I can obviously run this just directly through here. So, I'm just going to say run this. Give this a quick little ping. This is also going to give me a Google sheet URL. So, I'm just going to open that. Um, it's going to take a few minutes. It says between 5 to 15. That's just part of how the workflow works. So, the idea is everything will be populated in this document when it's done. Now, if you guys had a keen eye, you'll know that it said monitor progress in Slack earlier. Um, and you're probably wondering what the hell does monitor progress in Slack mean? Well, what I've done is I've set up on modal like uh service connector that dumps all of the thinking of the model into a slack channel called cloud-cloud. When you're working with agentic workflows, obviously it's very important that there's a level of interpretability here, right? Otherwise, models are just going to grow more and more blackbox and more blackbox. And so essentially what this does is this allows me to see how things are going in my workflow. It'll actually say, hey, I'm at step one of four right now scraping my query is dentist. My limit is 100. Step two, I'm uploading 100 leads to the sheet. Step three, I'm enriching emails with any mailinder. Enrich zero emails because I found all of them. Casualizing the names that even includes the information, right? Um this is particularly important when there's debugging occurring, right? If there's some form of debugging happening and you can see here I'm now identifying the debugging. there's some issue with the YouTube outlier flow. Like, I can see it. Otherwise, I'd just be sitting around waiting, sitting on my ass wondering when it's going to finish. So, this is beyond the purposes of this video for me to show how to do, but suffice to say, it is very straightforward. It's not complicated whatsoever. And maybe if you guys want it, um, a future video, I'll show you guys how to do it in like 5 minutes. Basically, anything is doable in 5 minutes now. But, um, yeah, very straightforward, simple way to gain a lot more interpretability into what's going on under the hood. Because if you don't stream the chain of thought I find or if you don't stream the steps the models or the executions or what step in the directives we are it can be pretty difficult to know what's going on whether or not you know a process is hanging because of a genuine wait time or because of some error. Okay, we did a little bit of back and forth with the agent just telling it you know which actor to use and whatnot. That took me a solid five or so minutes and at the end of it we received a YouTube outliers document that looks something like this. So we can see that we initiated this thing entirely autonomously or rather entirely um you know through modal we ran it and uh now we you know got the workflow set up to the point where we can actually grab both the thumbnails and then also a bunch of um information about you know what's going on. So we have like a highle summary and publish date and video links and so on and so forth. So I really like that thumbnail. Maybe doing some repurposing there. I like how there's some logos all over the place and the open AI thing goes to multiple things. Not as big of a fan of this one, but yeah, suffice to say, pretty cool, huh? So, that is how to do it on modal. All you really need to do to be abundantly clear is you just take the same document that I showed you, that cloud execution framework, and you just prepend it. And then, you know, ideally, you'd already have the directive orchestration execution framework set up. And then from there, you just ask it, hey, you know, can you take this thing which is occurring locally on my computer and put it on the cloud. Then, it can do a pretty good job uh most of the time. Hopefully, you guys appreciated that video. Had a lot of fun putting it together. If you guys have any questions about how agentic workflows work, how any of these cloud things work, or if you guys have more recommendations or suggestions surrounding um, you know, agentic workflow content, just drop them down below because I do take your recommendations into account. The reason I made this video was directly because of YouTube comment I got just a few days ago. Otherwise, thank you very much for checking this out. Definitely join Maker School if you want to achieve outstanding results in 90 days. We guarantee your first paying customer or your money back. Um, you can check that link at the top of the description as well. I'll catch all y'all in the next video.