Transcription
Hi everyone!
So in this video, I'll show you how to build a 20 AI agent team that can manage and automate literally almost any task or workflow across your entire tech stack. This is one of the craziest agent systems I've built up until today, and I think these kinds of setups will be the future of AI agents and automations.
This agent team has access to all my software, including all my communication channels: WhatsApp, LinkedIn, email, calendar, Slack, and even voice calling. It has access to my project management tools like my CRM, Notion, Google Docs, and Google Drive. It has three research agents to research any lead or topic, and it has specialized content agents that can write and publish across my social media and blog.
I can interact with my agent through voice messages on WhatsApp. The real strength of this system is that it can automate complex workflows across multiple software with a simple English sentence. For example, it can find flight options, add them to a Google Doc, and send it to someone on WhatsApp. It can call a friend to reschedule lunch and update my calendar with the outcome. It can research new leads that contacted me, add them to my CRM, and notify my team on Slack if they are qualified. It can write and publish blog and LinkedIn posts on the latest AI news, all from just a single request on WhatsApp.
These are just a few examples; the possibilities with this setup are really endless. In this video, I'll give you a demo and a full step-by-step breakdown of the setup, and the template will be available in my community. So stick with me because I think it will blow your mind!
Now, if you don't know me yet, I'm Ben. I've been building AI agents for businesses for over a year now. I also run a community where I teach others how to build and sell AI agents. If you're a business looking to adopt AI into your operations, you can also book us for a free call in the description below.
I'll first show you a demo, then I'll give you an overview of the setup for this system, and then I'll provide a detailed step-by-step breakdown so you can learn how to build something like this yourself. Remember, the template will be available in my community.
I'm just going to give you a few examples. Remember, these are just a few examples; the system is capable of a lot more, but I think it will give you a good idea.
Another cool thing about this system is that we can actually schedule messages for things we want our agent to do every day. For example, I instruct my agent to create a Google Doc every morning with all messages that came in across all my communication channels. With that document, I can start taking action with my agent.
For example purposes, I'll just instruct my agent manually now so you can see how it works. I can leave a voice message if I want. I can say something like, "Hey, please retrieve all unread messages from all my communication channels. Also, check all my meetings scheduled for this week and put them in a Google Doc and send them back to me."
Now, in the background, my director agent, who I call, is the one I'm in contact with. He is going to delegate these tasks to his sub-agents. To retrieve these unread messages from communication channels, he has a communication manager who has another six sub-agents.
The communication manager has a sub-agent for WhatsApp, LinkedIn, Slack, and all my communication channels. Basically, my direct agent will delegate this to the communication manager, who will then instruct all the other agents to retrieve the unread messages. That will be sent to my project manager, who has access to Google Docs, and will then put it in a Google Doc.
He will send that back to the director agent, who will then send it back through WhatsApp to me. As you can see, we get a little bit of a summary here on WhatsApp too, and at the end, we get the Google Doc.
Let me open it up quickly so you can see. We get an overview of all the unread messages from all my channels: WhatsApp, LinkedIn, Slack, unread emails, and my calendar events for this week.
Now, based on this document, I can start taking actions on these things, and my agent, of course, can start taking actions too. For example, my mom here asks, "When are you coming to Holland? Did you already book?" So you can see, actually, here's my mom's message.
We can tell our agent something like this: "Hi, please check flights from São Paulo to Amsterdam for the 19th of December. Check for the three cheapest options, put them in a Google Doc, and send them to my mom on WhatsApp. Ask her if these arrival times are good for her, and then also send the Google Doc back to me."
In this case, it's actually using the Google Ser API to check for Google flights. We then try to find the three cheapest options, again put it in Google Docs, and then I'll use the WhatsApp agent to actually send it to my mom.
We'll see it actually appear here, but it will send it to my mom from my WhatsApp. You can see it sent it to my mom now: "Hi Mom, I found three flight options from São Paulo, arriving at 11." You can see in the Google Doc we get the flight options too with the prices.
Of course, maybe not the best example, as I'm already in WhatsApp, but I just wanted to show you what the system is capable of.
We can also take actions on things unrelated to people, of course. I've also connected it with my social media and my website, for example. So we can also say something like, "Hey, please research the latest trends on AI coding agents. Write a blog post and a LinkedIn post about the latest news on AI coding agents. Post the blog post to my website and add the LinkedIn post to my content calendar on Notion, please."
In this case, it's going to delegate it to the content manager, who has specialized agents for writing blog posts and LinkedIn posts. If you've seen my other video, it basically has these fine-tuned post writers that can write optimized posts for different social media.
I'll actually show you that it will add a new article. So these are my latest articles, and you'll see that it will add one.
Alright, we got it back. The blog post on AI coding agents has been successfully written. Additionally, a LinkedIn post draft has been crafted and added to your content calendar. You can see it posted it. I think it made up the URL because it didn't get it back.
But here you go: "AI coding agents revolutionizing software development with landing capabilities. AI coding agents enhance software development with seamless integration."
Right, and it even added images, as you can see.
Let me check my Notion LinkedIn content calendar. Let me check: "AI coding agents revolutionizing software development." AI coding agents are about to change everything. Here are some of the latest developments: LLMs, new AI agents can write code to solve tasks and improve themselves over time. Super AGI is launching an AI agent that can code.
This is really good, actually! Auto GPT, so it gives a list. The coding process will be automated, and you can see it adds in my name and my CTAs.
Of course, we can also take actions on leads, which I normally do. For example, if we have two unread LinkedIn messages wanting to discuss creating agents for Sage Medic, I can go something like this: "Please research and scrape the LinkedIn of Christian App from Sage Medic. Add the enriched data and lead to my CRM and send Oscar a second message on the lead and the contact email."
Just an example, right? But you can see, you can basically just combine because you have an agent system that has all of these different software working together. We can sort of automate these workflows that require multiple steps, and I think that is really the powerful setup for this system.
I really think it's the future because we can basically start automating workflows with simple human language.
Alright, after that, we added Christian to your CRM with all enriched details. A Slack message has been sent to Oscar. You can view the contact details in the Google Doc, so it even puts it in the Google Doc with all the info.
Let me check Slack. Right, this is Oscar. You can see it comes from relevant Sage. "Oscar, we have a new lead," right, with all the information.
Now let's check contact. What's his name? Chris, was it?
Yeah, here we go: Christian App. Alright, we actually have the lead summary, as you can see here: company summary, company size, email, etc.
Well, you get the idea, right? Lots more possible.
I actually wanted to show you another one because these agents can actually call for me too. So I can literally tell my agent to call this person. I can literally let my agent call anyone, for example, to reschedule a meeting or an appointment, and it can then even update my calendar with the outcome of the call, for example.
But I think you got a good idea of it. Let me go over the system right now, where I think you get an even better idea of all the possibilities for this system.
As you can see, it's quite a big setup, but it looks more complicated than it is, and I'll break it down step by step so you get a very good idea of how this is set up.
Now, first of all, I've set up almost the entire system in Relevance AI, and I used Make.com to make some of the integrations we don't have inside of Relevance AI a little bit easier through Make.com.
If you're completely new to these platforms, Relevance AI is an AI agent and AI tool builder, and you can also build these really powerful multi-agent systems to automate quite complex workflows through these agent teams.
Make.com is a more traditional workflow automation tool with the big advantage of having lots of native integrations with third-party software.
Now, if you're completely new to this, this is a little bit more of a complex setup. I have many other tutorials on my YouTube channel too on both of these platforms, but I'm going to break it down in a detailed way and step by step. So even if you're completely new to this, you'll probably be able to follow this, and I highly recommend following it if you're interested in building no-code AI agents and AI agent teams.
So let me get through the system.
First, here on the top, we have the triggers. Right here, we have our director agent, how we call it, and the director agent is the one we are in contact with through WhatsApp.
Then we have our manager agents here, and here we have our sub-agents. Our sub-agents also have tools, and you can see here one manager agent also has three tools.
So in total, these are 20 AI agents and more than 50 tools.
You might be wondering why I use so many AI agents and so many different tools. Now, the main reason is because at the moment, unfortunately, LLMs and AI agents are not good yet at doing multiple tasks.
To get these systems where there are so many possibilities as reliable as possible, we want to reduce down the responsibilities and the tasks for each of these different agents and each of these different AI steps as much as possible to get the system as reliable as possible.
So instead of giving lots of responsibilities to one AI agent, we want to reduce it down as much as possible so each of our agents has very specific tasks to get the reliability higher.
So how does this work in practice?
Right, so here we have our WhatsApp trigger. Of course, that WhatsApp trigger, in this case, I send voice messages that go voice to text, and that will be sent to my director agent.
Now, I put these here too because that's the interesting thing with this system, I think, which is if we want to start automating things on a regular basis, for example, every morning retrieve all my unread messages, etc., we can just plan that in with your language.
This is why I think it's the future: because we can start automating workflows without actually programming in a workflow automation, etc. We can literally, with human language, schedule an automation that runs every day.
For example, you could say something like this: "Retrieve all unread messages from all my communication channels," and you can do that every morning at 7 a.m. That will be sent to my director agent, who will then perform that and send it to my WhatsApp or wherever else I want.
But you can also do other things, right? Research all new leads that contacted me through LinkedIn, email, add them to my CRM if they're qualified, send them a message back right away with a link, etc.
I think the amazing thing with this system is that as soon as you start interacting with this team and you see that you start doing things on a regular basis, then you could just build in a message to schedule every day to your agent, and you don't even have to manually trigger that anymore.
Now, from that, of course, it's being sent to our director agent. I'm going to go through this step later.
So a director agent already has four main responsibilities. The first one is breaking down our query. Then, second, according to the query, he'll have to plan out what he has to do: to which sub-agents or manager agents does he have to delegate these tasks?
Then he has to evaluate if the work that has been done by the manager agents and the sub-agents has actually been done correctly, and if not, send it back. Lastly, he has to communicate back with me.
That's why we gave our director agent access to two tools. The first one is to send a WhatsApp message so he can contact me, and the second one is to get the current date. Sometimes he will need to know what the current date is if I, for example, ask to retrieve messages from the last week or something like that.
Now, you might be wondering why we didn't just give this director agent access to all these 20 agents. Why do we put four manager agents who then control another layer of agents again?
It's because if we give our director agent access to 20 AI agents, it's overkill for one AI agent. It's going to make mistakes; there are too many responsibilities. You have to plan out and orchestrate a workflow automation across 20 AI agents, and that's going to go wrong.
Again, we want to limit down the responsibilities as much as possible, and that's why we have these four manager agents.
Of course, we have the communication manager agent, who has all the agents that control my communication channels. So we have the voice agent, the WhatsApp agent, the LinkedIn agent, the email agent, the calendar agent, and the Slack agent.
So what are the responsibilities of my communication manager agent?
The first one, again, is very similar orchestration and delegation. He will have to decide which of these sub-agents he has to use.
So in our example, to retrieve all unread messages from all communication channels, of course, he would have to use all of these sub-agents. But sometimes it might be only from one channel.
Then, second, again, he has to make sure that what these sub-agents have done is actually right, and if not, he can send it back and make sure that they do the right thing.
This is the second advantage of having these multi-layered agent systems: we can have these evaluation steps in there where agents can check what other agents have done and make sure that what they've done is right. If not, they can actually send it back until these sub-agents actually do it right.
Lastly, of course, he will have to communicate back to the director.
Now, if you're unsure about how this actually works, the delegation basically works like this: our director agent just prompts our manager agents on what to do.
For example, if I say retrieve all my unread messages from all my channels, the executive director will prompt the communication manager with something like, "Hey, please retrieve all unread messages for all Ben's communication channels," and then, for example, put it in a Google Doc.
Once the communication manager has done that, the director agent will then send that to the project manager, who says, "Please put this in a Google Doc."
So that's orchestration and delegation.
Then the second agent is the project manager agent. Same responsibilities again, which is orchestration and delegation, evaluation, making sure that what these sub-agents have done is correct, and of course, communicate back with the director agent.
Here we have the CRM agent, the Google Docs agent, and the Notion agent.
Then we have the research manager agent. Same responsibilities again: orchestration and delegation, evaluation, which is especially important for the research manager agent, where sometimes you want to do in-depth research.
This research manager can really double down on these sub-agents to make sure that they do in-depth research. If they haven't done proper research, he will send it back and make sure they do more.
Again, communication with the director agent.
The last one is the content manager, and the content manager has one more responsibility, which is that he has three tools: post to Webflow, post to LinkedIn, and post to Act.
He can actually post onto my social channels. Why do we do that? Because, for example, if our LinkedIn writer agent writes a LinkedIn post, we have another sort of check in place by sending it back to the content manager agent, who can check, "Okay, is everything good? Does it make sense? Does it match the original query?"
Then he can post it to LinkedIn.
As you can see, these sub-agents also have tools. These tools basically allow our agents to interact with software or even do more, like workflow automation.
In this case, you see our Slack agent has a tool where he can send Slack messages, retrieve Slack messages. We have our calendar agent who can get calendar events, create calendar events, update calendar events, and get calendar events.
I'm not going to go through all the tools, but you get the idea. Each of these agents basically controls a software inside of my tech stack and can perform actions inside of these software.
By having all of these agents with all of these actions in these different software, that's how we allow these agent systems to automate very complex workflows with human language.
That's why I think this is the future of AI agents. Because instead of mapping out a very rigid logic-based workflow like we used to in the automation world, and probably will rely on that for a while still, once the LLMs get better and more reliable, you can see that AI agents will be able to automate a workflow or process without building in this entire sort of logic-based workflow.
Now, what I'm actually building in right now, which will probably be available in the template too, is another extra step.
What I notice is sometimes, because this director agent already has quite a lot of responsibilities and some of these queries can be extremely complex, he has to delegate between four or five agents and make sure that he prompts each agent with the right thing.
For example, he can't prompt an agent with something he can't do. So breaking down a query and planning out is actually the hardest part, I think, or the hardest responsibility for this director agent.
That's why I'm trying to build in a GPT-4 model to actually do the planning for our director agent to take that responsibility out of his hands.
Because the GPT-4 models are, of course, very good at sort of system two-level thinking, breaking down and planning out.
So basically, I'm building a GPT-4 model because, unfortunately, we can't use them in agents yet, to first look at the query that I gave it.
I gave it all the context on the whole system, and the GPT-4 planner basically makes a detailed SOP on exactly what to do with this query for the director agent.
Instead of the director agent then actually having to plan before executing, he can just execute on the SOP and therefore reduce his responsibilities.
We can get this system more reliable. I'm building this in right now, so the system I showed you in the demo is without this step, but I think I will add that in the template.
Again, if you want to look at all the tools and all the agents in detail in your own time, the template will be available in my community, and I'll also make sure to put the link of this overview here in the description below.
Now let me get you through Relevance AI, where I'll go through step by step inside of Relevance AI how these agents are set up. I'll get you through some of the tools, and I'll also show you how to set up these triggers.
So here we are on my Relevance AI dashboard. Here we have some of the agents of this agent team.
Now, if you're going to clone this template, make sure to first make a Relevance AI account. I think sometimes it gives an error if you don't have an account yet.
So first, make an account. I'll make sure to put the link in the description below too.
Then, if you're going to clone my template, you can basically click the link and clone it, and then you'll see these agents appear in your own dashboard.
I'll first break down how this actually looks in the background so you can basically see what our director agent did with the queries I showed you in my demo.
So you get a good idea of how this system actually works sort of in the background.
Then I'll show you the setup for the director agent and the four manager agents quickly, and maybe some sub-agents and their tools.
Lastly, I'll also show you how I trigger the system from WhatsApp. I'm going to show you that very quickly because I have a full WhatsApp agents video too. If you're interested in that, make sure to check that one out.
Lastly, I'll also show you how you can schedule these repetitive messages on a daily basis, for example, to your director agent to automate your workflows.
So let's go to the director agent.
Here we can basically see what happened in the background. You can see here how I triggered it. For example, here: "Hey, please retrieve all unread messages from all my communication channels. Also, check all my meetings scheduled for this week and put them in a Google Doc and send it back to me."
Now you can see this was done with voice to text because I sent a voice message. That was triggered through Make.com, which I'm going to show you in a second.
Then in Make.com, we went voice to text, and then we sent it here. You can see the voice text isn't perfect, but it's good enough for our agent to understand.
Now we can see actually what our agent did in the background. You can see it performed four steps in the background.
The first thing it identified is that it has to use the communication manager. Here we can actually see what a director agent prompted to our communication manager on what to do.
You can see here: "Retrieve all unread messages from WhatsApp, LinkedIn, Slack, and email. Also, retrieve all scheduled meetings for this week from the calendar."
Now this information it already filled out itself because he has this context on what this communication manager can do inside of the prompt of the director agent, which I'll show you in a second.
Here we can see then what our communication manager did because he then delegated these tasks to his sub-agents.
You can see he delegated first to the WhatsApp agent, where he basically instructed him to retrieve the unread messages from WhatsApp. The same for LinkedIn, the same for Slack, the same for email, and also for the calendar agent.
He gets all of that information back from all these sub-agents, and then you can see this is what he sends back to our director agent.
Here, the retrieved unread messages for this week: WhatsApp, LinkedIn, etc.
The director agent receives that back, and of course, the director agent then has to actually make the Google Doc. That's why he has to use the project manager, right?
Because the project manager has a sub-agent, which is the Google Docs and Google Drive agent, who can actually make a Google Drive.
You can see create a Google Doc with the following details. He sends over all the unread messages.
Now the project manager delegates that again to the Google Drive and Google Docs agent, who has a tool to make a Google Doc.
You can see the Google Drive agent got prompted with the same prompt basically to create a Google Doc. You can see it created a Google Doc.
I can show you this quickly. You see all the text, and in the response, we got the Google Doc.
He sends that back to the project manager, who then sends it back to our director agent, who then, of course, sends us the WhatsApp with all the information.
You can see he sent us a quick summary of all the unread messages and, of course, the link to the Google Doc with all the messages.
That's sort of how it works in the background.
You can see the hardest thing for these agents is sort of breaking those queries down and orchestrating which tools and which sub-agents have to be used to perform this task.
But that's also the power of these agents. We can come with these very dynamic queries, and it can just know which sub-agent has access to which tools and then can sort of think through, "Okay, I first have to use him, then him to actually perform this entire workflow."
Now you can see here for the second one: "Hi, please check flights from São Paulo to Amsterdam for the 19th of December. Check for the three cheapest options, put them in a Google Doc, and then send them to my mom on WhatsApp and ask her if these arrival times are good for her."
In this case, it actually first has to do research on the flights. In this case, you'll see he first uses the research manager agent, who in his turn has access to two sub-agents.
You can see find the three cheapest flight options from São Paulo to Amsterdam. Include details such as airline, departure, arrival times, and prices.
You can see what a research manager did. He delegated it to one of his sub-agents, which is the travel agent.
The travel agent has access to tools to get flights, to get hotels, and things like that.
So you can see the travel agent, what he did in the background, right? He used the get flight option, found the three cheapest options, and sent that back to the research manager, who then again sends that back to the director agent.
Then he delegated, of course, again to the project manager agent to create a Google Doc out of it.
You can see we got a document. He delegated, of course, to the Google Docs agent. Lastly, of course, the WhatsApp has to be sent to my mom, right?
So it's delegated again to the communication manager agent, who can send WhatsApps.
You can see he sent the following message to Mom on WhatsApp: "Hi Mom, I found three flight options from São Paulo."
Now, how does he know what number my mom has? You can see here the WhatsApp agent actually has a tool to find the number from a name.
In this case, I call my mom in the database "Mom," but he has a database with names and phone numbers, so he can retrieve a phone number from a name.
You can see what he did here. He filled out the name and got back a phone number.
Then he used the send WhatsApp tool to actually send the WhatsApp to my mom.
Again, he sends that back to me, what he has done. I'm sorry, to the director agent, and he then sends all the information to me: "I found three flight options," right, and the document.
So that's sort of how it works in the background.
Now let me show you quickly the setup for this director agent. Here above, you have our agent settings.
Here you can see first we have our agent profile. Now the agent profile is not that important in this case; it's more important for if this is a sub-agent.
So here we have the agent name and here we have the agent description.
Now, why is this important for sub-agents? Because basically, the agent description will be read by the manager agent to know what this sub-agent can do.
Then we have another section here, which is the trigger section. We have lots of triggers we can use. You can see we actually have WhatsApp for business too here in Relevance AI.
Now, I didn't use it because with this one, we can't use voice interpretation, Google Docs interpretation, or image interpretation. That's why I set it up through Make.com, because then we can actually receive those types of messages too.
I'm going to show you that later.
So here then we have the agent instructions, right? The core instructions, which is basically our agent prompt.
Now, if you're new to this, agent prompting is a little bit different than normal prompting, and I always use my agent prompting tool, which I created myself.
It's not perfect, but it will help you write them a little bit faster, I think, and sometimes a little bit better.
The agent prompting tool is also available in my community.
So how's this agent prompt structured? Basically, first we have the role, which is always very important in these agent prompts.
"You're Ben's executive director agent responsible for overseeing and orchestrating the workflow of four manager agents."
Then I give it the names of the manager agents, which in turn manage a total of 15 specialized sub-agents.
Then we have the objective, which is sort of the high-level overview of what this agent has to do.
You can see his first one is delegation and orchestration: "Break down tasks by Ben down into subtasks and assign each subtask to the appropriate manager agents. Make sure you provide clear instructions to your sub-agents, but ensure that the tasks you're assigning to each sub-agent are actually doable by them."
Now this part is actually very important. The first time I created this system, that's actually where it went wrong a few times.
Because if you don't clearly say this, what sometimes can happen is that your director agent is going to prompt a task to a sub-agent which that sub-agent can actually not do.
So for example, in this example, I give like if you say retrieve all unread messages and add it to a Google Doc, it might say to the communication manager agent, "Retrieve all unread messages and put it in the Google Doc."
But of course, the communication manager can't make a Google Doc; that's the project manager agent.
In that case, something could happen where the communication manager says, "I can't do that, sorry," and the system breaks.
So it's very important that our director agent only prompts his sub-agents with tasks that they can actually perform.
That's why I have this, and that's why it's also very important to give a lot of context to this director agent on what all of these sub-agents can do and what they cannot do.
Now, the second thing here you can see is quality assurance: "Verify that the manager agents have executed tasks accurately and delivered outputs that align with the original instructions. If not, provide detailed feedback and request revisions until the outcomes are satisfactory."
The third, of course, is reporting to Ben: "Compile and send all relevant details of completed tasks and outcomes back to Ben via send WhatsApp to Ben tool."
That's his only tool, including entire message content, links, and results.
Now here I have an SOP where I basically break down these three tasks into even more steps so it knows exactly how to do this: delegation, quality assurance, and reporting.
I just break it down step by step: "Review the input from Ben and determine if the task involves multiple types of outputs or workflows that require collaboration across different manager agents."
If so, break down the task into subtasks, etc. Same for the quality assurance and the reporting to Ben.
This part is then very important, which is giving that context to this director agent of what these agents can actually do.
This is one of the most important parts in a system like this with so many agents: giving a lot of context on what these other agents can do.
You can see I give it, I instruct it very clearly what they can do: "Manages communication." I even describe which sub-agents our communication manager agent has, the key tools they have, and even an example task.
Now I do that for all the manager agents and then some instructions. The instructions are always good to sort of double down on important rules.
For example, "Use the send WhatsApp tool to deliver a deal to and compile a report to Ben."
Then we have some examples. Now the queries can be really dynamic in these agent systems, especially in an agent system like this, so it's hard to come with examples of input and output.
But it's still good to give an example of how it should approach a certain query.
In this case, I just give an example of a query and basically tell him what he would do or should do in that specific query.
You can see the action steps: "Assign the LinkedIn post task to the content manager," etc.
So he knows sort of gets context on what to do with a certain query.
Now this is exactly also what my agent prompting tool helps with. This will sort of generate it for you automatically if you fill out some other details.
You also have to double-check if it actually does it perfectly, but a lot of times it does, and it saves you a lot of time.
These examples really do enhance the performance of these agents.
I have a second example here, and then again the note section. This is to double down on important rules because, again, LLMs take instructions given to them in the end and in the beginning of the prompt more into account than instructions in the middle of the prompt.
So important rules always put them here. You can see I put in another one here because it struggled with that the first time I tried this: "It is vital to my career you only prompt your sub-agents with tasks that they can actually perform based on their capabilities."
Now we have the flow builder option too. Now, I didn't use that in this case because the flow builder is more if we have a very specific sort of set of actions our agent has to perform in a specific sequence.
Then the flow builder is really good to sort of double down on that.
In this case, of course, the variety of tasks and sort of orchestration can be very big, so we don't actually want to limit that through the flow builder.
That's why I didn't use it in this one.
Then we have the abilities here. We can label tasks if we want. It's just for here in the sidebar how it labels tasks.
Then we have actually an option of scheduling messages inside of Relevance AI.
Now, this is unfortunately a business plan feature, so I'm only on the team plan, so I also don't have access to this.
I'm going to show you a workaround how we can still schedule messages without this feature or without having the business plan.
Then we also have to escalate to humans, which can be very useful, especially for chatbots and things like that.
I show you an example also in the WhatsApp agents video. It can be very useful if you want to escalate it; you can escalate it through Slack or through email.
Then we have the tools. In this case, we only have one tool, which is the send WhatsApp to Ben tool, of course, to communicate back with me.
Here we have the sub-agent section, and here we have our four sub-agents.
Here we have some extra settings. The first one is the approval mode, so we can actually decide if we want to let this director agent run this communication manager agent automatically or if we want to have approval required.
Now, for some use cases, that can be very useful. When we send out a really important email or something, we always want to check before it actually sends it.
Then we can use these human-in-the-loop steps basically with the approval required.
Now, in this case, we want a completely automated system, so I put all the agents on auto-run.
Then here we have some extra settings, which are also important in these agent systems: prompt for how to use.
Here again, we give it more context on what this sub-agent can do. Very important for the director agent again to know what this agent can do.
So here we just double down on that again. We tell it what its responsibilities are, what it can do, the key tools it has access to, and an example task.
Then we have one more option, which is the template for communication.
Here we can basically decide what prompt our director agent should use when it communicates with the communication manager agent.
We can sort of decide that for them already.
We can say something in this case: "Retrieve all messages from," and then we can even use variables with the double curly brackets.
Those variables are basically prompts inside of a prompt.
Basically, this you will always have to send, and this is what we leave sort of open for our director agent to fill out himself.
Here we can describe to our director agent how he should fill out that variable.
So we could say the channels to retrieve info from or messages from.
Now, in this case, I didn't use the template for communication because you can imagine that if we use a template, of course, we also limit a little bit of the options of how to communicate with the communication manager.
In this case, there are so many options possible in this system that we don't really want to put these limitations on the system because it might limit the amount of workflows we can actually automate.
In this case, we didn't use it, but it can be very useful if you have a little bit more of a rigid system to get more reliability inside of the system.
Then we have the advanced settings. It's important to always use the best models for these agents.
In this case, I use GPT-4, but you can also choose other models.
That's it for the director agent.
Now let me get you through the four manager agents quickly.
So let me first show you the research manager.
The research manager basically has two sub-agents. I'm going to show you this.
You can see here in the agent profile the research manager. I put a quite big description because, again, this is what our director agent reads to really understand what this agent does.
So we have the core instructions. Again, role, objective, right? It's quite similar to the director agent in terms of his objectives, of course, because he also is a manager.
Then we give it some context and his sub-agents. We have the travel agent in this case and the research agent.
So I'm going to show you the travel agent in a second too and the research agent.
We give it an SOP too and some examples.
It's very similar structure to the director agent. Nothing really special here.
You can see we have the travel agent and the research agent and no tools available.
Now, important for the research manager agent, you can see also in the core instructions is to review the work that was submitted to ensure it matches the original task and requirements.
These are important tasks for these manager agents that are in between to make sure that what these sub-agents have done is actually right.
Because they're sort of in the middle between communication between the director and the sub-agent, and sometimes context gets lost.
That's why these manager agents, this role is also important for them.
I will show you quickly the travel agent, for example.
So you can see we have the travel agent. The travel agent actually has access to Google Ser API to do searches for Google flights and Google hotels.
You can see this is a lot more of a simpler prompt, as you can see, because this agent only has access to four tools.
It's a get airport code, get IATA airline code, get hotel option, and get flight options.
Now, these ones are to get specific codes for specific airlines, which are necessary to use when you call the Google Ser API.
Let me show you a quick example.
You can also see how these tools are set up for agent instructions. Here it's a pretty straightforward one.
As you can see, your goal is to assist Ben in planning his travel by searching for flight options using the required airport and airline codes and searching for hotel options, ensuring an efficient use of available tools.
So pretty straightforward.
You see here with a description of the tools and some examples.
If I just give the same example, like please look for flights from Amsterdam to São Paulo for, let's say, the 3rd of January 2025.
As you can see in the background, it first gets the airport codes of both because these are necessary to do the flight search for.
I noticed if I don't use these in a database, sometimes it actually puts in the wrong code, and then, of course, we can't do the search through the Google Ser API, which I'm going to show you in a second too how that's set up.
You can see we got it back here. Here are some flight options.
Now we can, of course, also specify this. We can, for example, ask it for which specific airline, and then we'll actually also use the IATA airline code tool.
These tools are basically databases, so I can show you quickly.
To get the airport code, what happens here in the background inside of the tool?
You can see if you're completely new to tools, tools are basically logic-based automations where we have an input field here, which we save in a variable, and then we can go through steps, logic-based steps to automate the process.
Of course, we can implement these AI steps in there too.
Now, in this case, it's a very simple one, only one knowledge search.
As you can see, what this is is basically a knowledge base of all the airport codes, where we have destination with the airport code, and basically, it just fills in this variable.
In this case, the city or country to find the airport codes for.
We have the destination, and it filled that out with Amsterdam.
Now that variable we use for the query of our database, and it goes to retrieve the most similar results.
In this case, I use search type keyword because that's the most efficient way to retrieve data for these specific types of searches.
Now, let's say you want to have a customer support agent that wants to retrieve an answer for a specific question.
That's where you probably won't use keyword but the vector.
You can see here vector.
Now, normally, the best way to do knowledge retrieval in Relevance AI is actually through the advanced knowledge retrieval, where we have some extra settings.
Instead of going either vector or keywords, we can actually go hybrid, which in my experience works best.
We can also choose the fields to vectorize, and we can do retrieval post-processing, which in this case is just a very simple search.
We want to go to Amsterdam, and that's directly in the database, so we can do this using the search type keywords to get an efficient outcome.
You can see if I run this, you can see we get Amsterdam back and the code.
You can see here I have the page size. Here you can basically decide how many results you want to retrieve.
In this case, this gets sent back to our agent so it knows the airport codes.
Now, in this case, we didn't specify an airline, but otherwise, it would use this similar tool to retrieve an airline code.
Then it's going to use Google Ser API to actually find flights.
You can see in this tool, we have the departure airport code, the arrival airport code, the flight date, of course, in a specific format, which are again specified in here.
If you don't notice, agents will actually read these descriptions here to know what to fill out.
So if you have specific formats, etc., make sure to mention them here in the description so your agent knows how to fill out these fields and what to fill out.
Then we have the return flight date. In this case, I didn't say that, and the flight class we can also decide: economy class, premium, business class, etc.
Then we have the airline code. In this case, we didn't specify it.
Now, it's important that you leave those options that are not always necessary or are not always given by the user as non-required.
Because otherwise, your agent's just going to make something up.
In this case, this one is not required, so it left it empty.
I also instructed that: leave empty if the user didn't specify a specific airline.
Right, amount of people to book for: use one if unspecified, which I do for me because I'm alone.
Then the two-letter country code of the country the user is visiting.
These are just sort of necessary to do the API call for the Ser API.
Now, I used Make.com to set up because Make.com already has the Ser API set up.
So I can show you that one quickly.
So there we go: find flight options.
Now, as you can see, we have the Ser API here.
If you don't know, the Ser API is basically directly from Google, and you can set it up here in the Ser API.
You can do actually some really interesting things with the Ser API.
As you can see, we can actually start getting data from Google Search API, Google Maps, Google Jobs, all these Google products we can use with the Google Ser API.
Now, in this case, I used the Google flights, but you can see there's lots of interesting use cases here: Google News, Google Trends.
So there's lots of interesting use cases with the Google Ser API.
Now, you can do 100 searches a month for free, and after that, you'd have to pay a monthly subscription.
You also have YouTube Search API, so very interesting use cases there.
Then you can set it up directly here. You get your API key, and it's pretty easy to do.
You add the API key to Make here, and then you have the Ser API.
You can see you can do all these things.
So how does this work?
If you're completely new to this, how do I make my agent sort of send information and retrieve information from this Make scenario?
It's pretty easy. All you can see here, right? We just use an API step here in Relevance.
If you're completely new to this, I have a full video on how to do this also on my YouTube channel, so I'll go through it very quickly here.
So you just click post, right? You use the webhook that you get from Make.
So first, in these scenarios, you create a custom webhook. In this case, I can't do it because I already have one, but you create a custom webhook.
You just look for webhooks, create a custom webhook, and in one click, you basically get a webhook.
So you copy that webhook, you create in this case here, right? Copy this, you go back to Relevance in the API call method post.
You post the URL in there, and then you can decide what you want to send over to Make.
Now, in this case, I sent over all these variables.
You can just add more here if you want, and here I put in the variables that our agent filled out.
So basically, make sure that all this information is separated and sent to Make.
Then for Make, we can then use it into the Google Ser API.
You can see in this case, I set up a router.
If the airline's not defined, or if the airline is defined, you can see I have a filter here: airline code is undefined.
Then it goes here because otherwise, it will give an error.
In this case, airline not defined one way, and this is back and there and back basically.
The same here, right? This is where the airline is defined, one way, and where the airline is defined, back and there and back.
So in Ser API, you can see all we need is the departure airport code, the arrival airport code, the outbound date, return date, and we map that to the values we get from the webhook.
Then Google will search for flights for that specific one.
Then we use an array aggregator to bundle all the results.
It actually gets lots of results, so I just choose the option of best flights, and then we send that back to Relevance AI with a webhook response.
We send back the array, the outcome of the array, the bundled outcome of the Google search API, and then we send that back to our agent.
So I can show you a quick example of how this would work.
You can see we get the flights back, and this flight options back, we got lots of flight options back that will be sent back to our agent.
Now, of course, in my example, my demo, I said choose the three cheapest ones.
Our agent actually reduces it down to the three cheapest ones and then sends that back to the manager agent.
So that's how this one works.
For the hotel options, very similar.
It's also with this Ser API.
I can show you very quickly, and I'll just do a search of the find hotels.
Now, in this case, a lot simpler, although we actually had to set this one up a little bit more manual, as you can see, because we didn't have the Google hotels option directly in the Google Ser API.
So a lot more straightforward, but similar process.
That's it for the travel agent.
Now let me show you very quickly the general research agent.
So here we have the research agent.
Now the research agent we have equipped with three tools.
We have a Google search, so you can basically do a Google search for any topic.
He can do web scraping, and he has a LinkedIn scraper.
Now we can add lots more if you want.
I have a full scraping agent tutorial also on my YouTube channel if you're interested in scraping other things.
In my specific use case, I only need this, but you can also set up social media scrapers, visual scrapers, anything you want here in your research agent, or even perplexity if you want things like that.
But it does a pretty good job with only these three because with the LinkedIn scraper, in this case, I'm going to show you through the example, right?
Search from my demo: "Search for Christian App's LinkedIn."
Using the Google search tool and then scrape his profile using the LinkedIn scraper tool.
Now, this was already instructed by the research manager agent on which tools to use because, of course, he has context on what this agent can do.
But you can see we didn't even have a LinkedIn from Christian, right?
He just used the Google search.
You can see Christian App even filled in the company name because he had that data point, and then LinkedIn.com.
He got back the Google search results for that and basically found his LinkedIn profile from those.
Then he used the LinkedIn scraper tool to scrape all of the data from his LinkedIn, as you can see.
He had lots of data, and then he made sort of a summary here of the details, which of course this thing can be used to actually, you see, email he found too, to actually update our CRM, which he did.
He even sent a LinkedIn message, I think.
But yeah, you can see this Google research agent is very important to actually do research on leads, but it can also research topics.
You can make this as fancy as you want.
In my other video, I have lots of other scraping tools available, which you can add to this agent too.
So that's it for the scraping agent.
I can show you very quickly in terms of the core instructions. It's not rocket science, right?
Using search tools, your goal is to assist Ben in conducting online research using the search tools to find relevant information and scrape useful data from websites or LinkedIn profiles, delivering detailed and clear reports to the communication manager.
So pretty straightforward.
I think that's it.
Then let me go through the next agent.
I think an interesting one is the communication agent, who has access to all my communication channels.
So let me start with the manager agent.
This manager agent, of course, has six sub-agents, as you can see: email agent, call agent, WhatsApp agent, calendar agent, Slack agent, and the LinkedIn comms agent.
Now, why did I call this comms? Because we actually have another LinkedIn agent who creates content.
For the core instructions, it's similar to the other manager agents.
We're in the middle, and we usually have three different responsibilities: again, delegating the task, quality assurance, and in this case, reporting back to the director agent.
Again, as always, give it lots of context on what the sub-agents can do and some examples.
So let me show you quickly the sub-agents, which I think can be interesting.
I'm going to skip over the email and calendar agent because I actually show those two in my personal assistant agent video.
If you're interested in those two, they're pretty simple.
Then we have, I think, an interesting one, which is the call agent.
He can actually call on my behalf, and it's a pretty simple agent.
Actually, all we have given him is two tools, right?
Which is, first of all, a database of my contacts, so it can actually retrieve phone numbers from names.
You can see tools: get phone number tool.
Now, this one's pretty straightforward. Unfortunately, I can't show you an example now because I'm actually recording with my phone.
But if you want to see an example, I show an example of the call agent or a sales agent that calls people in another video, which I'll also make sure to link up here.
So I'll show you quickly the tools, how to set up.
So get phone number: really easy.
Again, I just have a database, and I do a knowledge search.
So it fills out the name or phone number and gets back the results.
So I just use Google Contacts to download a list of all my contacts and put them in a database.
If you don't know in Relevance AI, it's pretty easy to set up a database.
Right here, knowledge. You could just create a table, upload a CSV, and that's how you create the knowledge base.
Then after you would, it will appear here in your knowledge sources.
So that's how this one works.
Pretty straightforward.
Then we have the call someone.
Now, the nice thing in Relevance is you actually have this sort of integrated, right?
Making the phone call.
Now, I think this is set up through VoIP, but we don't actually have to be inside of VoIP to set it up, which is the nice thing here.
The second nice thing is we can actually personalize the prompt, and that's exactly what we do here.
Because, of course, if I say, for example, "Reschedule a call. Call my friend. Ask him if we can reschedule lunch for Saturday at 1 p.m."
Every time my query will be different, so that's what in here in the input fields.
What I have is, of course, the phone number, which you'll normally use first to find the phone number.
You'll fill that out, and then depending on my query, it will fill out the goal and details for this call.
Describe in as much detail what should be done in this call, and the first name, of course, of the person to call.
Now we use this Make phone call step, where we, of course, the variable is the number.
We have the assistant system prompt: "You're Ben's personal assistant helping him call people in his network."
Here we have a pretty simple prompt, and of course, we put in the goal of above here in the objective.
So every time I give a different query or my voice agent is going to call with a different sort of script and different outcome, of course, and that he also understands why he's calling.
Of course, here I make sure that, you know, you will first greet the user, mention that Ben from Spendo asked me to reach out to you.
Now I can optimize this a bit more.
If you want to see more optimized prompts, also check out that other video, the sales multi-channel sales agent video.
That's it.
Then we have one more step here, where we actually retrieve the call details.
So basically, we got a transcript back.
We can even get the whole recording back if you want, but in this case, just get the transcript back because, of course, if I asked my friend if he could reschedule for Saturday at 12:00 p.m. and he said yes, then we actually also want to know that, of course, that was confirmed.
Then our agent can send that back that that was confirmed, and of course, then we can send it to the calendar agent to actually schedule in the meeting.
So that's it for the voice agent.
Then I can show you quickly the LinkedIn agent.
I think it's interesting, or the WhatsApp agent.
The WhatsApp agent has actually quite a few tools, right?
As you can see, we have six tools.
I'm going to show you why, and I'm going to show you these tools because I think there's lots of interesting use cases.
This is from my personal WhatsApp, as you saw, right?
It was sent directly from my own WhatsApp.
But yeah, again, if you're really interested in WhatsApp, I do have a full WhatsApp agent video too.
So basically, what we do with these tools here, and I'll show you through the example, right?
You can see this is the demo: "Retrieve all unread WhatsApp messages from today and report back with sender names."
So what it first does is get the unread WhatsApp chats.
Now I'm going to show you that very quickly.
It's basically just an option in the WhatsApp module, right?
So we get all chats, and basically, what you can see in the second step, I get all the chats back.
In the chats, basically means you just get the names of the people, the phone number, and if you have unread messages, basically.
Then I identify, right?
Extract all the JSON objects from the chats where unread count is more than zero, meaning he'll retrieve back all the messages or all the conversations that have an unread count in it, meaning all the unread chats.
But we can't actually, in this step, in this specific step, the get all chats, we don't get back the actual messages inside of the conversation yet.
That's why we have a second tool.
So it's a get unread WhatsApp messages tool, as you can see.
Basically, in the input field there, as you can see, we have the chat ID, which is also what that first tool brought back.
It's chat ID, and through that chat ID, we use the get conversation option, I think, in the WhatsApp module to get the conversations, to get the messages back from that specific conversation.
You can see it filled out the chat ID, and here we have get all messages from chat.
Then that, of course, is sent back to our agent again to get the actual conversations.
You can see it uses that for each of the unread chats we have.
It uses a different chat ID to retrieve the messages from each of the ones where we have unread messages.
Then it can also get the name from a chat ID.
Alright, so which is basically the database, right?
Then in the end, you can see we get the unread messages.
So a little bit of a setup, but it works quite well.
So that's it for the WhatsApp agent.
Of course, you can also send messages, right?
As you can see here, right?
Send a WhatsApp message to Mom, right?
So that message, right, where we have the same, we have the database to actually find the phone number for my mom.
Then it uses the send WhatsApp tool, where it needs the phone number and the message.
We use that same module, but then as the send a message option.
Again, this is not the official WhatsApp API, so you can literally do this with your personal WhatsApp number.
So you can see it can start new chats.
I usually use this one because then we can use the phone number.
We can also use the send message in the chat, but then we need the conversation ID, which would have been possible too.
But even if the chat has already started, you could still use this one.
Anyway, that's a little bit more technical, but that's the WhatsApp one.
Now for the LinkedIn one, very similar process.
I can show you. I'll just go through it very quickly.
So as you can see, we have the same thing: get unread chats first.
Then we have get messages from the conversation where we have the unread messages.
Then we have the send LinkedIn message, and we have one more here, which is send the LinkedIn invite.
In this case, we don't have a phone number, but in this case, we need to, for example, send a LinkedIn message or send a LinkedIn invite.
We need the LinkedIn URL.
So in this case, we actually let our research agent, who has the LinkedIn scraper or the Google search, find the LinkedIn profile before sending a message or an invite on LinkedIn, for example.
So here you can actually see it did it with that Christian App.
First, found his LinkedIn profile with the research agent.
That was then sent back to the comms manager agent, the LinkedIn profile.
He reported that to the LinkedIn comms agent to actually, you can see, right?
You instructed in with this, and you can see he used to send LinkedIn message to actually send this message.
In my demo, I didn't notice, but yes, I did it.
So it's good to also build the guardrails in, but it seems to be good.
He sent my sently out. That's what I asked him to, so good!
So yeah, that's it for the LinkedIn agent.
Now for the email and calendar agent, check out my personal assistant if you're interested.
Now let me go through to the project manager agent.
Here we have the project manager. Of course, the project manager, again, is a manager, right?
So very similar responsibilities as these other manager agents, right?
Of course, task delegation, orchestration, quality assurance, and reporting back to the director.
Of course, this one has three sub-agents: the HubSpot agent in this case, for my CRM, right?
Notion agent and a Google Drive and Docs agent.
Now I'll go over them very quickly.
So for the HubSpot agent, right?
Here we have the HubSpot agent.
Of course, what he can do in this case, I gave him a few tools, but we can do a lot more if we want.
If we want to expand on this, in this case, the most common use cases for me are add a contact to HubSpot, get a HubSpot contact, get information back for a contact, and update a contact inside of my HubSpot.
But you can add lots more actions inside of your HubSpot if you want, right?
Basically allowing your agent to do any task inside of your HubSpot almost.
So in this case, as you can see here, same example, right?
The research agent did all the research on this new lead, sent that back to the director, sent it to the project manager, of course, passed this on to the HubSpot agent, who then had the task to add Christian to HubSpot.
You can see add, use the tool add contact to HubSpot.
I'll show you very quickly.
So here we basically gave it lots of input fields if he can find extra data, etc.
We want to save that inside of our CRM.
You can see we have the email, the job title, LinkedIn URL, lead summary.
In this case, he doesn't have the company LinkedIn URL, so important again.
You can see I have these all on non-required.
Why? Because if that information was not available in the research, then of course it would error if he doesn't have the information.
But here you can see they did find the company site, so it filled that out.
So it fills out everything it can fill out, and then we just use the HubSpot API call here, which is a built-in module here in Relevance.
All you'd have to do if you use HubSpot is find the endpoint to find the path, which you can find in the HubSpot documentation.
In this case, the path for creating a new contact is this one, and the method is post.
Then all we use is this, where this is the property name inside of HubSpot, and this is the variable, of course, that our agent fills out.
Even if you don't know how to code, this is really simple.
That's it.
So it'll basically update our CRM and add that contact, right?
And that's what it did.
So that's it for the HubSpot.
Of course, it can do some other actions.
You can also get contacts, update contact, and if you want, you can add lots more capabilities to your agent too.
So that's it then for the Notion agent.
The Notion agent, of course, can have quite a few tools.
Actually, he has a tool to get my to-do list, to update my to-do list inside of Notion.
So these are basically databases inside of Notion.
Get my YouTube content calendar, update my YouTube content calendar, update my LinkedIn content calendar, get my LinkedIn content calendar, and it can also create a Notion page.
Now, Relevance AI doesn't have a native integration with Notion, so in this case, I use the same setup with Make.com.
I'm not going to show you all of them in detail on this one.
Again, blueprints will be available too for all the Make scenarios in the template.
That's what it can do.
You can see, right, in the background, same thing for the demo, right?
It, the LinkedIn content writer agent wrote the LinkedIn piece, right?
That was sent to the project manager, who then added it to my LinkedIn content calendar.
You can see, right?
And again, we use that Make.com to send it over to make it easier to update our Notion.
That's it for the Notion agent.
Then we have the Google Drive and Doc.
Now again, no native integration with Google Drive and Google Docs in Relevance AI.
Again, I kept this one simple, but you can add a lot more, right?
So in this case, you can create a Google Doc, right?
And you will actually save it right away to my drive too, and it can get Google Drive files.
Right?
So as you can see here in the example, again, right?
Same example, we got all that research information.
It also made a Google Doc, right?
As you can see, right?
So how does that work?
Right?
Same thing, right?
I send it over to Make to make the doc, right?
With the API call, so pretty straightforward, I think.
And one more thing here, which is HTML format, right?
To actually make it look sort of decent here, as you can see.
So that's it.
Then we have the API call where we send it to Make, where we have the Google Docs module already built in.
We send that text over there, put it in the Google Doc, and the Google Docs web view link is getting sent back.
So that's it.
Last manager agent is going to be their content.
I'm going to keep this one brief because I also have lots of videos on content agents, and I reused some actually.
So here we have the content manager.
Now the content manager has four sub-agents, right?
And actually, this manager agent is different from the others because he actually also has tools.
He has the options to actually post directly to LinkedIn, post to Webflow, my website, blog articles, and post to Act.
Now, why do we give that to the content manager agent?
It's because we get that double check.
We get the extra check, like I said before.
Instead of giving it directly to, let's say, the blog writer agent, we actually get the content manager agent to first check if everything's alright, that it matches sort of what we were looking for with the original query, and then he can decide to actually post it.
So we have those three tools.
Now again, I use Make.com for most of these to make these integrations easier with these platforms.
But if you want to know more in detail, I explain it in detail in my repurposing video.
A lot of referrals in this video to other videos, but it's because I reused quite a bit.
For the sub-agents, I'm not going to go through them all because I show it also in the repurposing agent video.
For this content agent, what I think is really interesting setup is giving them specialized fine-tuned models for each of the platforms.
AI, in my opinion, struggles the most with replicating a sort of natural tone of voice, especially for these specific platforms like LinkedIn.
So I give them fine-tuned models.
Now again, if you want to know more about fine-tuning, I have a full video on it, also how you fine-tune based on LinkedIn for yourself or for other people.
But you can, of course, do this for all different content types or social media.
That's, I can show you the LinkedIn one too, but I do have a full video on LinkedIn assistant.
So you can see here we gave the LinkedIn assistant a few tools.
It can actually also do some ideation, right?
So it can find similar posts.
Again, if you want to know this in detail, this is from the LinkedIn video, right?
So I can actually find similar posts from people I like in the space.
It can query a LinkedIn dataset to retrieve specific LinkedIn posts about specific topics.
Then I have my fine-tuned LinkedIn post writer tool that basically uses my fine-tuned LinkedIn post writer tool that writes it in a good tone of voice for LinkedIn.
He writes four variations, and then our agent system can actually choose which one they like best.
So that's how it works.
Now, lastly, I'm going to show you very quickly how you actually can schedule messages to your director agent so you can actually automate workflows on a repetitive or daily basis.
I'll show you very quickly how you can set up that WhatsApp trigger.
So we go back to Make.com.
That's the workaround, right?
We use Make.com to actually trigger this.
So basically, all we do is we create a new scenario, right?
Here we can create an API call to our Relevance AI agent.
Now, in this case, unfortunately, Make.com doesn't have any native integration with Relevance AI yet, so we have to set up an HTTP call, which I'm going to show you very quickly.
So we click here on HTTP, make a request.
There we go, make a request.
Here in the URL, if we go back to our agent, we go to API, and here we have an endpoint.
We're going to copy that endpoint and go back to Make and paste that in.
Then the method is always going to be post in this case.
Then we need two headers.
The first one is going to be content type, and you can also, I'll show you later where you can find this.
The value is going to be application/json.
Now I'm going to show you where you find this.
Here in the sample curl, you'll see header one, content type, application/json.
I actually put it wrong.
Alright, and the second header is going to be the authorization, and that's going to be your API key.
I'm not going to show you my API key, but you can generate it here by clicking on this button.
Your API key will be generated. You copy that and paste it in here.
Then you have your header set up.
Then you need to select the body type, which is going to be raw.
The content type is going to be JSON.
Then we need the request content.
It's like, what are we going to send?
So we go back here, and here we have the request body, and we're going to copy that and paste it in.
Now, as you can see here in the request body, we have the agent ID, and the agent ID basically tells Relevance AI to which agent we want to send this.
Now, we don't only want to send it to this agent; we also want to send it to the same conversation.
Because if we send it to the same agent but not to the same conversation, it will start a new chat, and it basically loses all the chat history.
Sometimes we want our agent, of course, to have chat history, especially in a system like this.
So we actually also want to add in a conversation ID.
So you can go in here, we can actually copy this part, right?
We add in a comma, and then we add in this part, agent ID, and then we change the agent ID to conversation ID.
That's it.
Then you can literally decide here the value.
Right?
So you could do one, two, three if you want.
And basically, this will tell the API call to always send it inside of this conversation, right? So that's it.
Now, in the message, we can define what we want to send. Let's say we want to schedule our agent to retrieve the unread messages every day from all my communication channels. Here in the content, this is literally the message we're going to send to our agents.
We can say, "Retrieve all unread messages from all my communication channels, put it in a Google Doc," etc. But we can say anything here, right? It's like, "Research all the new leads that came in, add them to my CRM, send a LinkedIn invite to them." We can do whatever we want, really.
That's the power here because once you've set this up, you can literally clone this and put in five different messages to send out every day. Now, once you've done this, that's it! You've set it up. All you do here then is schedule this.
So we're going to save this quickly, and now we can schedule when to run this API call. We can go at regular intervals, for example, every day at 8:40 a.m. This is going to run. So every day, it's going to send that message to our agent, and we'll receive that document with all the unread messages inside of our WhatsApp.
Of course, again, you can clone this and set up five or six. That's, I think, the power of this.
Now, let me show you very quickly the WhatsApp setup. This is going to be a little bit more complicated. If you want to know this in detail, check out my WhatsApp agent video, where I show you this setup in detail.
Here we have the integration with the WhatsApp Business Cloud. As you can see, it might look a little bit complicated, but basically, this does the API call that I just set up. This is the WhatsApp module.
So, this WhatsApp module gets triggered every time a message is received on this WhatsApp. If you don't have this, this is the WhatsApp Business API. If you want to set up the WhatsApp Business API but don't know how to do it, I also explain it in one of my other videos, which I'll make sure to link up here at the end of the video. I'll show you how to set up the WhatsApp Business API.
Now, if you don't want to go through that whole hassle and still want to be able to use this system, you can still do it. The only limitation will be that you can't do it with voice messages, documents, or images.
The way you would do that is you go to your agent here in the triggers and click on the premium triggers. You do need to be on the team plan, I think, in Ren evance, but you can click this and connect your own personal WhatsApp to this agent. You can trigger it through text messages from your own WhatsApp.
This is a possibility. If you want to set up the WhatsApp Business API, then you can have this system. Here, you basically have the router. All it does is these are media types.
So, this downloads the media types and saves them in Google Drive. Then, depending on if it's a voice message, it goes speech to text and sends it to our agent. If it's an image, it goes image to text or describes the image so our agent understands what has been sent.
Here's a document, and it will transcribe the document. Now, it might not be that useful for this specific setup, but transcribing documents and images from WhatsApp directly has a lot of use cases.
I literally know a startup here in São Paulo that just raised a lot of money. Their startup interprets documents and images, extracts information, and outputs it into a database. WhatsApp is used so much that this is actually a really powerful use case for many different businesses.
So, there are a lot of different use cases for this setup. If you want to know it in detail, again, the template I'll also put in the community. If you want to know it in detail, check out my WhatsApp agents video.
So, that's it for this video. Thank you so much for sticking with me if you're still watching. I didn't even go through everything yet, but again, if you want to check out everything in detail and really want to replicate it, you can join my community.
Of course, I'd love to see you there. Besides my templates, I also have one-on-one tech help and some other cool things in the community. If you're interested in taking building these systems seriously, I think you will like the community.
If not, that's fine too. I will keep making a lot of videos on YouTube anyway. If you got any value out of this, I highly appreciate a like, a comment, and a subscribe. I'd appreciate it a lot.
Let me know if you have any questions in the comments below. Thank you so much, and hope to see you in the next one!