Making n8n AI Agents Reliable (Human-in-the-Loop Demo)

If you've deployed any AI workflows into production, you probably know that human in the loop is a really good step to add in when you have things that are sort of publishing to social media or doing some sort of critical action. And inside of Nadn, for that very reason, we have some human in the loop capabilities for Slack and other apps. So today, we've got a really interesting story because this is an interview with a founder who's working on a dedicated human in the loop suite and their functionality goes deeper than NN's functionality. The reason I want to share that today is for a few reasons. The company is a solo bootstrap founder team that's built a rather impressive offering. They also built an NDN community node and got it verified. So they're now discoverable on NNN cloud and growing through NN as a distribution channel. So I wanted you to hear his unfiltered opinion on that experience. And then we have a demo of the human loop capabilities to also show that Naden is really about our community. So if you build something that overlaps with some functionality we have if we see that it's useful what you're building and that it's a value ad on top of Noden we're more than happy to scream about you from the rooftops because we love our user base and we love people who use it in and improve it by making a robust ecosystem where you as a user have choice. So on that — so on that note let's meet till Simon

the founder of go to human. — How did you first learn about an event? Maybe someone told me about it at a meetup or so. I knew a competitor and then I was very excited when I saw there was a new player especially coming from Berlin and I thought it was amazing especially since had a slightly different focus allowing to have some code in it to be more flexible and all of that and seeing that it has that had that momentum because when I saw it was starting to see that momentum. I'm the founder of go to human. Go to human as a human in the loop solution that you can integrate in your AI workflows to supervise and orchestrate them. So let's say you need a human approval somewhere in your workflow. You can use go to human for that. — What is human in the loop? Assuming I just entered the tech bubble. I've got no idea what is it and why is it important. Basically means if you have your workflow and it's running and you have one step where like well here we need to include a human get in the loop you know and provide input. and why it's important. I think that's a very interesting question, especially right now. There's the thing with AI workflows, right? You see all these magical demos online, especially on LinkedIn, and they seem to be these agents that do everything automatically with just minimal input and instructions, and you can see some great results. But the thing is, as soon as companies think about putting something live, like rolling it out in production, then you need predictability and reliability. And then even if it's just 10% of the cases that your workflow goes off the rails, that's a big problem and it's really hard to fix, right? What a lot of companies do is first of all, they try to minimize the AI parts to not have just one big AI agent, but have more deterministic steps in there. But what a lot of teams do is actually include a human in the loop because then you get reliability. you have an expert that's included and then provide input and can verify that there's no errors, no mistakes, no hallucinations or even like the wrong tone of voice. And that's how teams often times get to reliable outputs by having that human in the loop. — What was that genesis point that aha moment you're like, okay, I've got to start a company that tries to help solve this problem. — I ran into this problem in an earlier project. I was using AI and LLMs and I was desperately trying to get this to full automation. I was working a lot on this trying to amp it up the last 10 or 20%. This has been a while ago, so wasn't even as good as today. At some point, I was talking to people about that old project. What kept coming up is just like the only thing that really works is if you consider the AI output being drafts, but then have someone involved and have that human in the loop. So, I started research, is there anything I can use and just plug into my project to not having to deal with that review system? And there was nothing there that was built for the AI era. So I started building first prototype. I rebuilt it probably two or three times already. — Was the name of your company AI generated? — No, honestly not. I remember it's the thing about these names that now I like it a lot because I'm getting positive feedback. But funnily enough, I remember I asked AI about it obviously, but the suggestions were really mediocre. Let's say it's like one field where maybe it's better now, but where it doesn't work super well, I would say. — Good question. It's a human engineered company name. — Exactly. Well, you launched Glow to Human and you decided to create an integration in N. N also has some human in the loop capabilities. So, can you walk me through why you decided to make the integration and how you see these two different functionalities and how they interplay — in Nad there are already human in the loop options and they work well for simpler cases and we will see later when we look at the product how go to human is built up and there are some things around it that are a bit more advanced and gives you more options towards it. I was lacking these kind of things and that's why it made sense to also build that integration. — You built the integration for end yourself, right? Uh as a verified community node. Is that right? — I built the integration myself and then yeah got it verified by the team. — What was that experience like? Walk me through it. — Pretty nice. I think what's cool is that you can code it to begin with because it could be done in different ways, right? But it's nice that you can code it so you can use your regular tools like git and do proper versioning and all of that, right? And it gives you flexibility obviously within the code. And then you also offer even two different ways to code your node and declaratively and programmatically right so that's nice also to have the option of complexity I would say there were some challenges in building it but also because go to human has two specifics to it that probably a lot of other nodes wouldn't have there's this dynamic node interface like the fields that you see they're dynamic there's web hooks involved in the background so the node might pause actually for hours or days until it continues it's amazing that nn has that functional ity that people might know from your send and wait human and loop nodes that are built in. But that's super great to have that. — How long about did it take you to implement that? — I think a couple of days. Yeah. — Any tips for folks wanting to build their own community notes. — Most important part is look at the other ones and that's the great part about it. It's open source so you can look at the nodes. You can then look at nodes where you know that they are similar. So I obviously knew which are the ones that have send and wait functionality and those web hooks. could look at that code and that's amazing. And we talked about how long it takes. I think if you have nothing too overly complex with a regular API, I think you can do it within a day, especially because you have those references that you can look at. And then in my case, normally people wouldn't need to do that, but I actually sometimes actually checked out the code of NN itself and because it's open source, I could have a look at it. So that was amazing and it helped me reason about one or two aspects that I wasn't sure about and got it to work that way. So that was amazing. — What did you prepare for us today? What are we about to look at? We're looking at a demo workflow that's AIdriven. Obviously, it's a marketing example, but I think it's going to be interesting because it involves images and also video generation. So, that's going to be cool. And it's actually generating a marketing campaign post. Yeah, it's set up to be on a weekly schedule. — Very cool. Let's do that.

— We're here in Go to Human. Till, take it away. Walk us through it. — All right. Let's have a look. So, we're in the Go dashboard. This is the back end of it. So you can see here, this is where you set up the approval steps and connect it to your agents. We can actually have a look at one. I already set this up for our demo. Let's look at a simple one. If we go to the details here, you can see already this has been running a bunch of times already. You can see the data that's been going back and forth. But more interestingly, let's look at the actual review template. The way that this works is that you can add different types of fields. This is a very simple one. You can see a preview here on the right, and you can see this includes a list of checkboxes. Now I could go in here and actually define for each checkbox what's the option, what it's labeled and so on. But in this case we're actually having dynamic checkboxes. So we are populating that from our workflow run. — What other kinds of fields did people get to play with when they're building this? — Let's have a look here. So you can see there's different options. Text based options obviously numbers, images. We'll see actually that the image here supports video as well. Markdown is very nice for like blog posts and stuff like that or just research reports, JSON, but also if you review emails that are supposed to be sent out. That's so nice. Editing JSON in plain text is just like the worst nightmare. — Absolutely. We can have a quick look here because you see it in the preview, right? You can edit it right here where it's shown. So that's pretty nice. We got some different options here to make the experience even better. Different user input fields that you can add as well the regular forms, right? Like checkboxes, drop downs, etc. We can jump over and you'll see this is the inbox that reviewers will find items here that need attention that need input. So they will come up here. Actually quickly jump back once. I forgot one thing to mention here. You're also included a manual trigger. So that's also an option that you can build a let's say trigger form that you can use to actually kick off your workflow. Some sometimes you want that manual option to be included. It will appear up here as a button right now. And this one was set up to simply take a campaign name in this case, right? So it's a very simple one, but that's an option to do that. Cool. All right, let's jump over to our Net end workflow. First part that's going to run is the ideation part. We actually want the AI to come up with ideas. Hey, what could be nice topics or for that image that we want to post or video clip? It's doing it mainly on seasonality. So, I'm passing it the date because obviously the AI will not know what day and time it is. But let's get a flower wheel. Okay, and we'll see how it works. The agent is working right now to come up with a list of topics. And now it went to go to human and it's in the waiting state right now. Let's jump over. It appeared here. Actually, we get like a nice campaign name to make it simple. It's based on the current date and topic ideas is the current review step. — Cool. So this is what the AI came up and I can pick from it. Let's see if there's anything that we like. So let's just go with this one. We'll actually in the next step have the ability to edit this again. So this one kicked on. You can already see it continued now in the adjust campaign. So this is for mainly for generating a tagline as well. So we want some little tagline on our post. — And we can see this ran. Now again we're in a waiting step on the go to human. — Exactly. Look again. So the tagline is in there but also like we repeat the audience here. I'll actually change that because it always works nice. Change the setting a bit so I can edit it here. — As someone that's tried to build hacky versions of this basically with like Slack bots, I can say I'm thoroughly impressed by the breath already of your app. — I picked the style that I want to have. So, this is linked to a prompt that's later on used. The scenario is we are an online ceramics shop background here. — Ceramics studio. — Very fast on your feet there. ceramic studio — hypothetical products in here. We'll just pick one of them because we want to have a regular post with different product. Let's pick you can choose. — Ooh, what do we pick, gang? To be honest, I'm liking the ombre fade here on the right there. Let's do that. — Let's roll. All right. And down here is our tagline. Crafted urban elegance. — So, this was the AI generated from that agentic step right before this. — Absolutely. Let's say at least we want to see an alternative, right? So, I clicked retry waiting state. We'll see. We can actually jump over it. — Is it going back into the workflow? Okay. So, — it went into the loop over here. You maybe quickly saw it running, but it's already back because this one is a quick one, right? — Got you for the folks at home. So, what happened is this outputed and then in this switch node, I guess this outputs some metadata on what that decision was. You check that, you route down here. You're resetting some input data to agent, right? Seen this pattern before we see agents rerun and then it reruns that back here. So, it's nice clean loop. We're not duplicating our agent multiple times. They just I've done that one before on a khaki one. — We've all been there. Yeah. — And then it actually updated it as you can see here. — See the version. — Exactly. What's always important I think with these human and loop approaches is what I mentioned earlier to be able to navigate also within these AI iterations because you don't know what's coming out. Right. The worst thing is you do this and then you're like actually the first try was the best and this is four or five words, right? But imagine you have like you have a full paragraph. — I also commend your AI agent for using a dash not a EM dash. It's really high-end fonting right there. — Let's say we are good with that one and we'll disapprove. So, all right, workflow continues. You can see it running. It went down here and we are here in our image generation. So, we're using file AI and it's using the flux model in the background. It already managed to output that image. We call it reference image because we're going to use it later on as well. And here we go. Here's the product you picked with the tagline. We can quickly go back to the step because here we got the workflow navigation. Yeah, — nice. So we can see. Yeah, that's the one that you picked, right? Jump over here. Looks pretty cool already. There's one thing that I don't like too much is that everything is sitting on the floor. — So I open the chat window here. What I can do here is actually edit the prompt because from the workflow I passed along the prompt. And now I can edit it. Now I will just edit up here saying we're standing on let's say a wooden table and I think somewhere down here says surface. Let's just emphasize this. It's a table. What I've done in the past just looks nice is let's say city rooftop with blurred out skyline. — So let's say send — you can already see like it went to two out of two — and is it also in the editend workflow cuz that was going to be my question. I wasn't sure if your AI capabilities then you just go to an LLM where you can use the exact same setup where your web is running. This is far I think more preferred in the use case. — Yeah. I mean because the main reason behind build a human is that you can plug it into whatever way you build your workflow, right? Is that the control over how you prompt and so on should be with like the person that built the workflow, right? So careful to stick with that concept over here. You can see that here's a new option. We can quickly go back. Yeah, that was the old one. This is the new one. And also within the prompt, we can still do like a retry, right? Goes pretty quickly. So I'll do that. Maybe you're familiar with this a little bit. like cloud and chat GPT similar navigation structure of going back and forth. — But that's a super important step to be able to navigate within these kind of chat like trees all you got different branches — and I think it's one of these things from all the like home roll solutions that have done of this cuz I'm sure a lot of people watching this have done their own home rolling of this stuff cuz we've all had this pain point. They're usually stateless cuz that statefulness is more effortful and then you really feel it. You're like, "Oh, why did I click it again? It was better the last time. It's not perfect but I would have picked that. " — Absolutely nice. And this part can get quite complex if you play it through, but I think I like this one. I don't know. You can say which one you prefer. — I think that's a winner. a winner. — Let's go with a nice creamy background. — All right. So, that was version three and we're good. Let's continue. Let's see what our workflow does over here. Continues down here to our video generation. We could already post it, right? It looks pretty nice, but let's make it more vivid. — I'm sure of the video capabilities going. Could be. Very cool. I'm thoroughly impressed. This takes some seconds, right? This might take half a minute or something. — That's the nice thing is I've seen these kind of workflows. You basically have to sit around and wait for it. You're doing it in a way where you can't do the long wait and stuff. These are async processes. It's great. I'll go to my meeting. I'll have my coffee. I'll come back. There could be 15 of these or maybe my colleague can go knock out those or vice versa. — And this is generating like a 5-second clip, right? Imagine like in the future it will for video it will be capable of generating minutes of it, right? So let's have a look. Review clip. What makes sense always for this demo I didn't add it is to have like a little plant in there that would that were kind of shaping the wind but what it does is like you saw a little insect flying around some clouds — but already this is going to perform better than the static image — just to make it a bit more vivid. — I think it's one of these things if I was a marketing manager if I had to go write to Vlad to go animate this from the image and then wait for Vlad for 2 3 days maybe he's got a problem with a home and this and then he writes back and there's a problem and back again. I'm not going to do this. It's not worth it in that case. But if it's incrementally, I don't know how many cents that cost, even if it's a couple dollars. Plus, I can just click click, then there's a whole world of possibilities of things that I would not have done before that now I'm thinking, well, let's do it. — Exactly. Think about it. If that was a small ceramic shop, they would just not do it. — Exactly. This example is perfect because the first one wasn't ideal. If you said, hey, look, the LMS are getting in 95% 98%. That's my brand. The 2% I can't have a terrible post go out of a mangled ceramic pot. every single time we go to check it becomes a lot easier to do so. Oh, — nice. So, a little tree came out. — Very cool. Very cool. cool. — I mean, it might even if the model is perfect, it might stick to your prom very closely, but maybe you know it sparks new ideas or there's something that you didn't think about. Remember the table like it didn't in the prom didn't never said that it needs to be on the table. So, sometimes it's about ah right. Yeah, maybe I should define that. But the model was right, but I just didn't specify it. So, sometimes you need the iteration to actually be right in your specification, right? Could we have a look in Oh, I guess we've got a — Yeah, we can then prove that. — This is good to go. Yeah. — One thing to mention maybe is also these are images and videos. They are stored on our CDN. So, they are cached within Go to Human and that matters because if you deal with something for example use image generation from OpenAI or some other services often times these links expire. So, what might happen if you send that link maybe you send it to your own Slack channel, you do that manually, the link might have expired and then it's not accessible anymore for your review. let alone for even processing that further right so good even makes sure that those are — they're signed links some of these images sometimes have like the o on them and the person reviewing this you don't want them to have access to your open AI or your zor or whatever but you want them to have a look at an image and click a button you can kind of separate that out — exactly that's still being taken care of but in the background so these are unguessable links — can you walk me through the sort of wait and approve steps — we can have a quick look let's look at a very simple one first here you can see first of all obviously you set up your credentials you find your IP the I key in go to human and this the component is the image grid. This is a fairly new capability right now. So the video works within the image grid. And what you pass along is a list of URLs. Let's say you could label that as well, but we're just passing a URL and this URL is coming from the previous step. — And then for that workflow navigation, we're actually putting something into this metadata field where we have specific fields that you can define to say, hey, this is the workflow that this belongs to and this is the previous step because the workflow can branch and so on so this can support it. Is that how we're getting the nice like titles and stuff for the different step that we're in on the approval? — Yes, exactly. Even here you see the update for review ID because we need to know if you're coming back from loop and there's an existing one. So that's one example — and I see what's this assigned users. You walk me through that option. — Yeah, down here you can also pick whether you just want to put the post a review to your account so it appears for everyone in their go to human inbox or you can actually add specific email addresses of people that you want to assign it to. So cool. So this inbox is kind of shared by default. So it's like you got a pool of people. Hey, anyone on the marketing team can approve this or anyone customer support can approve this AI response to customer. — Let's have a look maybe at an earlier one where there was a bit more happening not the image that will look similar here. It's probably a bit more happen in the product style. We pass we have these different fields like the audience scenery and activity that I actually edited I think right to the rooftop party. And this where we pass along the products. I made this dynamic also considering that maybe you have some dynamic source of where you get your products in here. I simply passed along the image URLs because obviously that's all you need. But what you could imagine if you were extending this workflow is grabbing like more data from your product data set, right? And then have maybe links where can you purchase that in your shop because it would make sense and then posting to have a link. — Yeah. Shopify step. It's pulling perhaps the filtered or something was pulling at this point. A little data manipulation beforehand and whatnot. You can pipe it in here for — Yeah, exactly. And then tagline is in here is just where you can see how we passed along the prompt and this is how it was available in go to human. — If I may, for example, I'm a new user. I'm like, wow, this is really cool. I want to set this up. If we just maybe expand this one. So here you've got an array of objects have a URL makes sense for the image grid. How do users know this structure — in go to human? You can have a look in the docs. There's also some hints to it but in go to human for your specific review template you can click on API request and you'll see the structure of it. Beautiful. — So if you go to this specific example you see API request. — Oh great okay so — and here yeah it open HTTP like an N it links to the guide but here you can also see API keys up there and also you see the structure. So for product you could see I could pass a long English array and we omitted the label but you can also pass a label and sometimes the node just expects text. So you can just input text and then you don't worry too much about the format for certain fields you have to have a look up for to be entered. — That's go to human and n very cool. I'm thoroughly impressed. This is an authentic reaction. No one's paying no one here. It's just kind of showing some cool tech. Um tell this is a pretty cool use case. I think it also is a really nice demo. Very artfully done. Shows the different functionality of product. Is this going to be a free workflow template for people? — Yeah, of course. We're going to share it. — Okay. We're going to get them to upload this for everyone cuz every time someone doesn't do a guest, you guys all complain in the comments. So we're going to get tilt to get this uploaded. So that'll be available for people and then go to human. If I like it as a user, how do I like is there a free trial and how does that all work? — Exactly. There's even a free tier through a certain amount of reviews. You can just use it for free and then yeah, you can see if you need more pro plan. — There's a few internal use cases in edit. I think I'm going to chat with the team. we're going to have to try this out cuz it's really robust what you've built out and I love tools like this was with Naden was always my ethos is I'd rather have a little bit of complexity and assume that my users if they're motivated they're going to they'll learn this but give them the means to make that last mile themselves. I think that's why and then successful in one sense because we don't have that glass ceiling and I think your product in a very similar way I think my brain is still just like figuring out that the possibilities of this will tilt. Thank you so much for contributing this. building community node and spending some time to share this with everyone. I think everyone's going to find this really interesting. I really appreciate it. — Thanks for having me. — My pleasure. And I'm going to hold you to uploading that workbook template. All right, we'll do that. — Happy programming, everyone.