How to automate anything with Multi-Agent Systems (no code)

In this video, I'm going to be teaching you how you can make multi-agent systems so you can connect up all of your AI agents and have them controlled by one single agent. Using this method, you can do pretty much anything you like because you can have multiple different agents that specialize in different things and can use different tools, all controlled by one agent that you can talk to. If you haven't seen me before, my name is Rory Ridgers and I'm an AI enthusiast, and I'm passionate about helping business owners integrate AI into their business. So if that's something you might be interested in, then I've left a link to my website in the description and you can go on there and book in a call with me and we can talk about some ways to integrate AI into your business. Also, just want to mention that all the templates that I use in this video will be available for completely free in the description. If you go in the description, you should see the link and you can download all of them for free. So let's not waste any more time, let's get into the video. So I've put together a little mind map so you can sort of visualize what we're going to be making in this video, and also so that you can sort of visualize how the structure of these multi-agent systems look if you've never seen them before. So it all starts with a main assistant here. I've called it Personal Assistant 2.0 because if you've watched one of my previous videos where I made a personal assistant. This is kind of, kind of be a more powerful version of that. So put it Personal Assistant 2.0, and you can see it connects to these 3 agents here. We have the data retrieval agent, the calendar management agent, and the email assistant agent. So when a user inputs something, it will go to the Personal Assistant 2.0, and then they'll choose what agent they'll need to use to be able to sort of fulfill the request of the user, whatever the user said. So then each of these agents have their own separate tools. So the data retrieval agent has the Search Airtable Data and Relevance AI Data Retrieval Tool. We'll come onto those later. We'll have a look inside of them. And then the Calendar Management Agent, you can see the Google Calendar Event Retriever and then the Create Google Calendar Event. And then the Email Assistant Agent has the Email Generator Tool and the Email Retriever Tool. These sub-agents here will get an input from the Personal Assistant. They will then use whichever tool they see fit, get the output, and then pass it back to these Personal Assistant 2.0. and then that will be given to us as the user. So this is kind of the structure of how these multi-agent systems set up, and this is what we're going to be building in this video. So if we come over to RelevantSR, you can see the agents here: the Personal Assistant 2.0, the Data Retrieval Agent, the Calendar Management Agent, and the Email Assistant Agent. So I'm just going to give you a brief sort of look into how all of these agents work and some of their tools. We're going to start with the Email Assistant just because it's by far the most simple. And I'm not actually going to go into the tools for this agent because I did make this agent in another video. I'll link that video in the description so you can go watch me make this, but the tools are really, really simple and it's a really, really simple agent. I'll just give you a quick breakdown of how these tools work and then also the prompt for the agent. Okay, so we have the email generator tool and all this tool does is generates an email and send it. So the input will be something like, send an email to Rory saying 'Hey, what's up? Are you still up for that meeting?' Anything like that. And basically all it does is it looks for a knowledge base and tries to find the email for— in this case it'd be Rory— tries to find that email address and then just use an LLM step to actually craft an email body and an email subject line and then just sends it all off. So that's all that does. And then we have the email retriever tool, and all that tool does is it will get input saying something like, 'Has anyone emailed me about the project?' or has anyone from this email address emailed me recently, for instance. And then all that happens inside the tool is the last 100 emails are returned, and then we just use an LLM step to go through that data and try and find what it was the user's looking for. So in this case, it would look through the data of the last 100 emails and try and find anything about the project, or try and find anything from this particular email address. So relatively simple tools. As I said, I'll link the video where I actually build these, so if you want to see how these tools work, then you can go watch that. Then we can come over to the core instructions. Pretty simple. Again, the role: you're a helpful email assistant that can send emails for a user and also retrieve data on emails that have been sent to them. And the task: your job is to recognize whether a user wants to send an email or retrieve data on emails, and then use the appropriate tool to fulfill that request. You have two tools: the email retriever and the email generator tool. Here's a breakdown of the tools. And then when I'm breaking down the tools for these agents, it kind of— I structure it the same for all of them. I say what the tool does to give it a little bit of context, the inputs. So the inputs are email request and email recipient. Sometimes I give an example, I don't always, but I also tell the agent how to input the data into these tools. So that's the email assistant. As I said, pretty simple. We don't really need to look too much into that one, but you can go watch a video where I make it. Now we're going to look at this calendar management agent, and this agent is able to create events on your Google Calendar, and then it's also able to retrieve events from your Google Calendar. So if you said to it something like, hey, I've got a meeting tomorrow at 2 with da da da, then it would create an event on your Google Calendar. Or if you said something to it like, what meetings have I got coming up in the next week, then it used the Google Calendar event retriever tool and it will come back to you with all the events that you've got in the next week. So I'm going to show you inside of these tools because they're a little bit more complicated than the email ones, and I think one of them also has a Make.com scenario in it. So I want to show you that. So we'll look at the create Google Calendar event one first. So the user inputs for this are event title, start date and time, end date and time, and then the attendees. So, so the event title, you'll see later on, sometimes you don't have to input this. If you don't say a title for the event when you're talking to the AI agent, you just say something like, add an event to my calendar for tomorrow, then it will just set the default to meeting, I think it is, but we'll have a look later. And then start date and time. Also show you how that sort of works in a second. And then attendees, this would just be whoever it is you're having this meeting or this calendar event with. So the first step is going to be retrieve data from knowledge, and this knowledge base here is an email list with names, as it says. And all it is is just some demo data that I put together. I can show you it. It's basically just a long list of names with email addresses on them. And oh, you can see it's a PDF, so it's not actually in the table. So this PDF here. We can open this and then you can see, yeah, this sort of looks like— this is all just demo data, these aren't real people. The only email in here that's actually valid is my one. So what we do in the tool is we just return that full PDF data and then we use an LLM step to basically look at the attendees here, because you don't have to always type out the email address of the attendee, you just want to say their name. So what we do we return all of this data and then we use an LLM step to go through that data and look for the email address that matches the attendee name that we've inputted. So if I said a meeting with Rory, it would then look through that data, find Rory, find my email address, and then output that. We save that as attendee email. And then here we have Python code and all this does is it just returns today's date. I think I made this tool a few weeks ago. I think since then Relevance AI has made a tool where you can, um, just return today's date. But this works perfectly fine. It just returns today's date. And then we need this because we need a reference point for if you just say to it tomorrow, because you don't want to say to the AI agent, um, schedule an event for, and then the date if it's just tomorrow. You just want to be able to say tomorrow. So that's why we do that. And then we basically here use an LLM step to take the output from here, look at the start date and time that the user said. So that would be tomorrow if you just said tomorrow. And then we're basically formatting the data correctly because we need to format it like this with year, month, date, and then the time— hour, minute, second. And then for me it's +100 because of my time zone. You can see here, give an example. For example, if the input is today at 5 PM, then output this date and then 5 PM. So it just needs this Python code here to return today's date to give it a reference point and then it will just format it correctly. And then we have another LLM step here that it will basically just take the output from here, which is the start time, and create the end time. Sometimes when you're talking to the agent and you're using this tool, you don't always put an end time. You might just say tomorrow at 5, and then by default it's going to just add an hour on unless you say otherwise. So if you say tomorrow at 6 until 8 PM, And obviously our output's 6 PM and 8 PM as the start and end time. But if you don't say anything, it's just gonna add an hour. And then yeah, we just format it the exact same. And then we just use this create Google Calendar event and basically just input all the outputs of those LLM steps into here. So event title, I just remembered actually that it isn't in the tool stage here that it creates an event title if you don't state one. It's actually in the AI agent step. So before it even comes to this tool, if you don't state a title, it will just pass the information to the tool as meeting. The AI agent will just input the event title input as meeting if you don't state a title for this event. So put the event title in there, the start date and time, it's going to be the output from those LLM steps, the end date and time, the output from the LLM steps, and then the attendee list is going to be the output from that first LLM step. They've got the email address and then it just creates the Google Calendar event. So not too complicated. It looks a little complicated with the time thing. That's kind of the most complicated bit, trying to get it to format it correctly, but it works every time when you use these prompts for the LLM steps. So that's perfect. And then we can look at the Google Calendar Event Retriever tool. And as I said, this one will retrieve events from your calendar. So let's have a look inside of it. So the only input here is start date and time, because the way I set this tool up is you can do it for the day. So you can say, what events do I have today? What events do I have tomorrow? What events do I have on the 25th? Whatever it is. So the AI agent will input the start date and time here, and this can just look like tomorrow or something like that, or today, or just the date. And then again, we use the Python code to get today's date as a reference point. Then we have an LLM step here, and it's kind of doing what it was doing in the other tool where it's formatting it correctly like this. So if it's Google Calendar events for tomorrow that we're looking for, we'll get the today's date, and then this will format tomorrow's date in this format. And then what we do is we output 24 hours on from that, so it outputs two separate variables, which will be— if we say, what events do I have tomorrow, this would be tomorrow's date here, and then this would be the day after dates here. That way we get the full day And then what we do is we have an API step here, and this is linked to a Make.com scenario. So I'm going to show you that quickly. So this is how the Make.com scenario looks like. Basically what's happening is we have a webhook, so that connects the Make.com scenario with Relevance AI. And then we have this set multiple variable step here because when we send data over to this scenario, we're sending the start and end time which is this. So it looks like this. So it's one variable, but we need to make it two separate variables. So what we do, if you get this LLM step to output it with a comma here as the separator for the two variables, then we send that over to the Make.com scenario, and then we use this equation here. So these get and split equations, what we do is we put in the value from the webhook, which is going to be this, the two separate variables. And then what we're basically saying here is we're splitting it by the comma. So that's the point in which the variables are split, is a comma and then a space. You have to make sure you have the space in if there is a space. And then this 1 just says we want to get the first variable here, and that's going to be the start time. And then the end time is going to be the exact same thing but number 2 because we want the second variable after that comma. So then this step will output 2 separate variables, which will be the start time and the end time. You can see we labeled them start time and end time. This will make a little bit more sense when we actually run this tool through, so don't worry if if it's kind of going over your head at the moment. Then we just use this Google Calendar search events, connected up my Google Calendar account. I've done the start date and then the end date, and then updated from— I've just put in the start time. Single events, yes, doesn't really matter. Nothing in the query. And then event types, all of it. And then the limit I've done to 10, but you can obviously set this to whatever you like. Depends how many sort of events you have on your Google Calendar. And then we have an array aggregator. So if you don't know what this does, basically it takes loads of different variables and it kind of bunches them together as one array. And that's what we want in this case to send back to the Relevance AI agent. So we choose the variables that we want— is just total number of bundles, the description, the summary, the start and end time, and the attendees. Because you can see all of these different variables that we get here, we don't want that. That's just going to clutter up our Relevance AI tool. It's going to make the LLM's job a lot harder after. So we just select the variables that we want. That'll create an array that we then use a webhook response to send this back to the webhook, which is obviously triggered inside the Relevance AI tool. So we're just sending it back to the Relevance AI tool. You can see here the array. And then in the Relevance AI tool, we use an LLM step to basically just summarize what is found. So if there's nothing, then it will just say there's nothing. But if there is events, then it will just output these events and give the user sort of a bit of a summary— who's attending, what the meeting's about, that kind of thing. So we're just going to give this a run through really quickly. So I've just added these events into my Google Calendar. We've got meeting with client at 1, and then gym, and then dinner with friends. So we're going to try and see if this tool is able to return this to us. So if we just put in tomorrow, which is when these events are, and then hit run tool we should see this return to us those events that we've just put into the calendar. Okay, so this is what's come back to us with. Here's a summary of your events for December 2nd, which is tomorrow: meeting with client, gym, dinner with friends. Enjoy your day filled with productive meetings and fun times. So let's just quickly look through what happened here. So this Python code, it looks like it says null, but it's the s out here. Today's date, da da da. LLM step is then outputted today's date and then 24 hours on from that. And then we triggered this Make.com scenario and we can have a look at this. So this is how the data came through. Then we split up as two separate variables here. So these are the two variables that we got. And then we used a Google Calendar step here and this is what we inputted. So it's done 1st December at almost midnight and then end date 24 hours on. And then you can see this is all the data that's outputted. Lots of variables here that we don't really need, so that's why we use this array aggregator step and we output this array here. So a bundle like this, just 3 nice bundles like that, and then we give that back to the AI agent via a webhook response. So it comes back to this API step here, this is what it looks like, and then we use an LLM step to just go through and summarize that. So that's how that tool works. So we can now have a look at the instructions for this agent. So the role: you're an assistant that can manage my Google Calendar. You can create calendar events and also retrieve events from my calendar. Task: you'll be triggered by another agent and it will be a request to either create a calendar event or retrieve a calendar event. I want you to take these inputs and use the right tool for to fulfill the request. Here is a breakdown of the two tools that you have, and then you can see these breakdowns of the tools are a little bit meatier than the other ones. I've said to it, this tool will create Google Calendar events. So just saying what the tool does. And then the reason these a bit bigger is because the data needs to be inputted in a proper way. So the inputs are, uh, event title, start date and time, end date and time, and then attendees. Use this tool when a user says they want to schedule a meeting or create an event or something similar. And then this is where I've said, if a user doesn't specify a name for the event or the event title or description of the event, then as a default use meeting in the event title input. But if they do specify a name or description, then obviously use that in that input. When a user specifies a time, make sure to fill out both the start date and end time input with the full date and time. For example, if the user says schedule, schedule a meeting for me and Sally at 5 PM tomorrow for an hour, then in the start date and time put tomorrow at 5 PM and then as the user said hour, put tomorrow at 6 PM. And then for the attendees input, just put the name of the person that the user specified they're having this event with. And then the Google Calendar event retriever tool— this tool retrieves event from the user's calendar for them. The only inputs for the tool are start date plus time. I want you to use this tool when you just say something like, what meetings do I have next week, or what is in my calendar for tomorrow. I want you to take the date and/or time of the user's input and use that in the start date plus time input for the tool, and then given an example. So that's what that agent does. You can create events on your calendar and also retrieve events from your calendar. Now we're gonna look at the last agent we have, which is the data retrieval agent. We can hit edit, and this agent, as you can see, is able to retrieve data from an Airtable database or from Relevance AI documents. So I'm sort of gonna show you why I've done it that way in a sec when we look at the tools. So we have the Relevance AI data retrieval tool and then also the search Airtable tool. And the reason I've done these two separate tools is because the Airtable tool, basically what we're gonna be doing is I've set up an Airtable database that has just a bunch of client information. And you could also use a CRM for this. It's kind of the same thing, but I thought I'd use Airtable just because you can do a lot more on Airtable than a CRM. So I thought it'd be better to show you how to connect it to Airtable. And yeah, like I said, we have lots of client information on there, so when a user says to this data retrieval agent something about wanting to retrieve some data on clients, then it will search the Airtable database. So if a user says something like, uh, what is the status of this project, for example, then it will search the Airtable database and find that and output it back. And then I've also got the Relevance AI data retrieval tool. And what this is, is I've uploaded a bunch of fake sort of company documents to the Relevance AI database. These are things like procedures, best practices, uh, the type of font that a company uses, that kind of thing. I've uploaded that to Relevance AI. So if a user wants to get some data on something like that, like as I said, what type of font do we use, or what's the procedure when this happens, or what's our client onboarding process, then it can search the Relevance AI data to find that information. I basically just split up client data for a company and then just every other kind of data. So I'm going to show you inside both of these tools. We'll start with the Relevance AI data retrieval tool because it's probably the more simple one out of the two. So the inputs here are user query, and this is going to be what they want to find. So what data? So like I said, what kind of font do we use? What's the process for this? That kind of thing. That's going to go in the user query input. And then what we do, we use a retrieve data from knowledge step. So this just returns all the data from this table. I'm going to show you inside this table quickly. I've just put together 5 example documents and I've pretended that I'm a web design agency. So you can see these are the docs we've got: Design Style Guide Template, uh, SEO Best Practices for Web Beginners, Website Launch Checklist, Client Onboarding Checklist Document, and then Common Client FAQs. And then these here, these are just PDFs with sort of a demo data in there. So what happens is all of that data is returned. So all the data from the PDF, not just the links to the PDF, all that data is returned. And then we use an LLM step to take that data and then also take the user query and just output what the user was looking for, basically. So we're just going to try it out quickly. I'm going to say, what is our client onboarding pro Success. So I've said to where, what is our client onboarding process? You can see this is what it's outputted. I'm not going to click view all because there's a lot of data there. And then the LLM step has just gone through that, all of that data, and tried to find the client onboarding process, which it looks like it's done. So initial consultation, contract and agreement, assess collection, project roadmap. So perfect. Yeah, it looks like it's done that perfectly. So this is obviously great when you don't want to look through all the documents that your company has to find something specific. You can just ask this AI agent, how do I do this? Or what's the process for this? What's the procedure for this? And it'll just search through all your documents and find what you're looking for, essentially. So now we can look at the other tool, which is Search Airtable Database. And there is another Make.com scenario for this, so I'll show you that. But firstly, again, the only user input is user query. So this would be something like when a user says what's the update on this client, or how many clients are we doing this particular thing for, that kind of thing. So that would go into there. Then we have an API step. So this connects to a Make.com scenario, which I'll show you. So this is how the scenario looks. We've got a webhook here, and then we have an Airtable search records here. I'm going to show you inside of that. So I've connected up this demo database that I made. It's sort of like a CRM, but as I said, I just wanted to show you Airtable as opposed to CRM, because I think Airtable is a bit more flexible with the amount of data you can put in there. So I thought it'd be more valuable for you guys to know how to connect to Airtable. So you can see we've got names, business names, notes, status, and then the website URL. Pretty basic, kind of what you'd see on a CRM. So what we do is we search through— I've added my base business data, table 1, grid view. We search through this data, we output all of these fields I've set the limit to 20 just because I think I've only got— yeah, I've only got 10 things in there, but obviously you'd set this to however much you want. And then what we do, we're using an array aggregator again just to output what we want. So the name, the notes, the status, the website URL, and the business name. We get that as an array which we can then send back to the Relevance AI agent via a webhook response. That comes back to the AI agent and we use an LLM step here. That basically just looks through all of that data and then takes the user's query and tries to find what they want. So basically what we was doing with the tool before with the Relevance AI docs, it's just looking through its data, taking the user's query and trying to find what it is they want. So we're gonna test this out quickly. You can see in here some of the statuses are leads. So we're gonna ask the agent, what leads do we have right now? So what leads do we have right now? Let's hit run tool. The API step going off. You can see that happens pretty quick. This is what's gone on. So we've got the data here, it's gone through the Airtable database, outputted all of this, put it all into an array that looks like this. So all these different little bundles here with just the information that we actually want, send it back to the AI agent, and then the LLM step has given us this. So name, business name, notes, website URL. The— these are obviously all the leads that we have here. So we're just going to check that out quickly. So John Doe, Lisa Morgan, and Tom Harris. John Doe, um, Lisa Morgan, and then Tom Harris. So yeah, perfect. That's exactly what we want. We can basically ask her anything about this data. So that is really, really cool. Okay, so now we're going to look at the main attraction here, the Personal Assistant 2.0 that's going to be able to talk to all of these agents. So The name, Personal Assistant 2.0. In description, you can communicate with other agents to be able to send and retrieve emails, create and retrieve calendar events, and search databases. And then of course, rather than going to tools this time, we're going to be going to subagents, and then we'll add in our agents that we want. So the data retrieval agent, email assistant agent, and the calendar management agent. And then we can look at the core instructions. The role, you're a helpful assistant that can communicate with other agents to be able to send and retrieve emails. Create and retrieve calendar events and search databases. And then the task, I want you to take user's input and determine their intent and then send that query to the correct agent to fulfill the request. You have 3 agents that you can communicate with: the data retrieval agent, email assistant agent, and calendar management agent. And I've said I'll give you a breakdown of each of these agents now. So the email assistant agent, these descriptions here are really quite simple compared to the tool descriptions because It doesn't need to input the data in a certain way. It can just send the data to these agents as if we was talking directly to the agents, if you know what I mean. So it can just be normal plaintext inputs. Doesn't have to be any kind of special inputs to go into tools. It's just to the agent. So just said what the agent does. This agent is able to send and retrieve emails from my inbox. I want you to trigger this agent when a user shows intent for wanting to either send something or retrieve data from that email. Take the user's input to you and input that directly to this agent when you trigger it. So basically all I said for all of the agents is what the agent does, when it should use the agent, and then how it should input the data, which in this case is just as it gets it, input that directly into these agents. So you can see for some of these as well, give an example. So this could be, what meetings do I have tomorrow, or what is in my calendar for next week when I'm talking about the calendar management agent. And then for the data retrieval agent, also gave some examples of what documents we might have for the retrieve data from Relevance AI tool. So these documents will contain things like procedures, best practices, company info, more, just to give it a little bit of context and make sure that it's sending the data to the right agents every time. So now we can test this agent out. So we're going to start simple, say something like send an email to Rory saying, uh, you— so send an email to Rory saying, are you up for that meeting? And it should talk to the email assistant here. Let's have a look. Yeah, so the email assistant, we can view the conversation as well, which is quite cool. So this conversation I had with the email assistant, you can see he used the email generator tool and then the email assistant outputted the email has been sent to Rory. So then let's look what the Personal Assistant 2.0 said. Yeah, it's just outputted what the Email Assistant said. The email has been sent to Rory with the message, are you up for the meeting? And then if we look on our emails, meeting availability coordination. Hey, are you up for the meeting? Let me know if you're available and we can coordinate a time that works for both of us. Look forward to hearing from you. So that's perfect. That works absolutely fine. So now I've said to it, do I have a meeting with Darren tomorrow? Just see what it does. I don't have a meeting with Darren tomorrow. And this is what it outputted back to me. It used the tool Calendar Management Agent, which is perfect, and it said, you don't have any meetings scheduled with Darren tomorrow. Here's a summary of your events for December 2nd. So meeting with a client, gym, and dinner with friends. So that's perfect. Hasn't outputted just my meetings for tomorrow. It's also recognized what I was actually asking it, which is, do I have any meetings with Darren? And it said to me, no, but you do have these events, which is really, really cool. And it's also recognized that it needs to look through my calendar to be able to know if I have a meeting with Darren tomorrow. So that's really cool. Now we're going to say something simple to it, like, so I've said, what font do we normally use? And obviously it's going to have to search through the database here, so it should use the data retrieval agent. Okay, so it is the data retrieval agent, and it should come back to us with the font that we normally use. Okay, so in the company's documents, we typically use the following fonts headings, the body text Open Sans. Cool, perfect. We're going to try and get some data on some clients now. I've said to it, what clients have an interest in emphasizing SEO for their sites? Because obviously we're pretending that we're a web design agency here. So should search through the Airtable database using the data retrieval agent. Perfect. It should come back to us if there's any clients that have an emphasis on using SEO for their sites. So Yeah, this is what it said. Here's the information regarding clients' interest in emphasizing SEO. So the name Emily Green, business name, the notes needs SEO optimization and blog integration for existing websites. So perfect, that's exactly what we want. It's managed to search through this data and find kind of a weirdly specific question, what clients have an interest in emphasizing SEO for their sites. So yeah, it's perfect, it's proven that it can handle those kind of requests. Okay, so you should now know exactly how to make multi-agent systems that can do pretty much anything. Obviously in this video I had it so it can manage your emails, your calendar, and also retrieve data for you. But any AI agents that you make, you can just add it to one AI agent and communicate with it through there. So it's kind of infinite what you can do with this. Any existing AI agents that you have as well, you can link all of them up in these multi-agent systems, or you can make new ones. So there's a lot of potential with these multi-agent systems, and they're definitely a great thing to implement into businesses. So speaking of that, as I said at the beginning of the video, if you do want to go ahead and book in a call with me, we can talk about some ways to integrate AI systems like this into your business. Then I've left the link to my website in the description, and you can go ahead and book in a call with me. And also, if you wanted to get the template for this video, so that's going to be all of the agents in this system, it's also going to be all the fake documents that I put together, I'll put them in as well, and also all the prompts for all the agents and all the tools and everything like that. All be available for completely free in the description if you just look at the link that says free templates. So thank you for making it to the end of the video. I really, really appreciate if you guys subscribed, and if you have any suggestions for videos that you'd like to see in the future, please do just drop them in the comments of this video.