Full transcript
0:00In this video, I want to talk about a
0:02strategy for building agents in N8N
0:04that'll allow you to build bigger, more
0:06complex bots that don't necessarily
0:09break, that don't take forever to run,
0:11that actually work much smoother, and
0:13aren't going to destroy your computer.
0:15So, this bot right here is actually one
0:17that I built uh when I was first
0:20starting out. Like, I I had no idea what
0:22I was doing. I just had this massive
0:23idea of building like this giant
0:26organizational flowchart and like I had
0:28a manager on the top and then you know
0:30submanagers under that. And so my
0:32thinking was well I need my submanager
0:35to be able to handle just all of these
0:37different kinds of processes and tasks
0:40that I would then be able to send to
0:42other agents. And what I have found was
0:46uh my computer couldn't handle it. My
0:48computer actually started to uh to melt
0:50just a little bit. Uh because when you
0:52have this many nodes and you have this
0:53many processes and this many different
0:55you know things that are running um
0:58whenever you hit test workflow my
0:59computer would just break. It just it
1:01wouldn't even run. In fact there came a
1:02point where I could I couldn't even add
1:04new nodes to this uh because anytime any
1:07amount of data would come through it
1:08would just just lock up and seize. And
1:10so I since learned how to better manage
1:12my processes where I'm able to use what
1:15we have these execute workflow nodes
1:17here that are able to call other bots in
1:21my account that those bots don't have to
1:24be active. I don't have to worry about
1:26them being live and using one of my
1:28live, you know, uh parts of my account
1:31that I'm only I'm limited to 15 on the
1:33the pro uh pro plan, I think it is. Um,
1:37and so, you know, you want to make sure
1:39that you're thinking through how to
1:41build these bots in a way that allows
1:43you to make a bot work, right? That
1:47doesn't melt your computer and is
1:49actually going to be better suited for
1:52processes down the road as you're
1:54building more. Now, this bot over here,
1:56this one that I have here, that it had
1:58one purpose, and that was basically to
2:00take ideas, turn them into different
2:04action items, and then process them in a
2:06way that would then be sent to various
2:09other managers and agents down the line.
2:12And I never got to that part. I was
2:14never able to build out that part
2:16because there was just so much
2:18complexity to it that it just didn't
2:20work. And instead I ended up having to
2:23do like I said um to scale it back have
2:26just a front end that then sent to
2:28different bots. And so what I want to do
2:30is really quick I want to show you just
2:32kind of um top level how to build
2:34agents. All right. So just we're going
2:37to do a click trigger. I'm going to
2:38build the most simple agent in the
2:40world. Here we go. AI agent. We're going
2:42to do um define below. We're just going
2:44to say hello and you're friendly. That's
2:48it. System prompt. I'm not even going to
2:49change the system prompt. Just do a
2:51quick chat GBT. Do a 4.1 nano so it
2:54doesn't cost me any money. 4.1 nano. All
2:58right. Um I don't need that. Um and then
3:03that's it. So most simple agent in the
3:06world. All right. We click test. It's
3:09going to output a response. Bam. Here's
3:12my response. Hello. How can I assist you
3:13today? That is technically an agent. All
3:16right. It has an input. It has an
3:18output. Now the output I'm not
3:20necessarily sending anywhere. I could,
3:22you know, do a Google Sheets sheets and
3:25I can add it to a sheet, maybe update a
3:27row, um, you know, whatever, whatever I
3:30need to process it. But this is just
3:32kind of a top level. This agent is doing
3:34the work. And I can give it a tool. I
3:36can give it like a Google calendar tool.
3:39So that when I instruct it to do
3:42something and I would have to set up my
3:43credentials and all that, like a
3:45calendar, you know, I don't want that
3:47one. I want this one. Uh, this one, that
3:50one, just so it gets rid of the error.
3:51Um, so I can give it instructions on to
3:54how to use this tool. Now I can load up
3:56other tools. I can load up drive. All
3:59right, we'll add that. I can load up um
4:01here. Delete the Google Sheets. I can
4:02add the Google Sheets here. Sheets. I I
4:05can add all these tools to this one
4:06agent here. All right. And at some
4:09point, roughly, I'd say five tools.
4:12Let's do Gmail. Add another one. Gmail.
4:15All right. Uh that one. Okay. And then
4:18we'll do one more. Say tasks. Do I have
4:21tasks? Well, tasks. Yeah. All right. So,
4:25once you've reached kind of a certain
4:27threshold of tools, the agents start
4:30having issues. They're not necessarily
4:33able to use those tools as well as you
4:36would like. And so instead of building
4:39just one agent and I I see a lot of
4:41creators on YouTube doing this where
4:42they'll add like 10 tools under one bot
4:46and I'm sitting there like geez like
4:48that's that's going to have issues at
4:50some point especially if you're trying
4:51to get it to do multiple things. Like
4:53don't try to get an agent to do multiple
4:55things. They're just not built to do
4:57that yet. All right. Well, we may get
4:59there, but like if you're going to have
5:02multiple processes, so I need my agent
5:05to pull all informations from Google
5:07Sheets. Um, well, for starters, do you
5:10actually need the AI agent or could you
5:12just use a Google Sheets node? That's a
5:13good question to ask yourself. Um, I
5:16think a lot of people are are using
5:18agents because it's the hot thing to do
5:20when you're just wasting tokens. Um, so
5:22that's just my opinion, but say I need
5:24it to to pull something from Google
5:26Sheets and then I need it to process
5:28that data and then send it through an
5:30email to something. Well, I guess you
5:32could technically have that all done in
5:35one agent, but you're now you're you're
5:37you're going to confuse the agent at
5:39some point because there's just too many
5:41instructions. You know, I have to do
5:43Google Sheets and then I have to send an
5:44email and then I need to add to Google
5:45Task and now I need to do it to this.
5:48You know, agents are much better when
5:49you string them together. So, I'll do
5:51another AI agent. Do another one. All
5:54right. Just do a define below. All
5:56right. And then we'll just add it to the
5:58same chat model. Now, the the 4.1 Nano
6:01is not going to be able to use a tool.
6:02So, that was a bad example. 4.1 Mini can
6:04use tools. Uh but here, let's this
6:07Google tasks. We just bring it over
6:08here. Make this guy do this task
6:10instead. All right. And so by by having
6:14your agents
6:16compartmentalized, you're more likely
6:18going to be able to have your task
6:20completed correctly for one. Uh but two,
6:25um it's also going to bring down your
6:27costs because you're not going to have
6:29to have these monster system prompts
6:32that have to be run multiple times.
6:34Instead, you're going to have a smaller
6:36system prompt that runs maybe just once.
6:38So just say for sake of argument for
6:40just Gmail. All right. And then your
6:42next one does just Google Tasks. And
6:44then your next one would do just Google
6:46Calendar. You're gonna have a much
6:47smaller system prompt. Your token cost
6:49is going to go way down. You're going to
6:51save money. All right. Now, the other
6:54thing that you can do, and I would
6:55actually recommend doing something like
6:57this. So, um, you can make this whole
6:59thing an agent or a tool itself. So, not
7:02an agent, you would make it a tool. So,
7:04we'll do an when executed by and we're
7:07hook this up. So now I can have another
7:10agent somewhere else in another process
7:13call these agents as tools. And this is
7:16what I recommend because this will
7:18really allow you to scale up whatever it
7:21is you're doing. Every single type of
7:23task that you might need done, whether
7:25it's getting emails or doing something
7:27on a calendar or pulling information
7:29from error table or what have you. You
7:33can set up the inputs however you want
7:34using this one executed. So, um, I just
7:37add a field, add a field, add a field.
7:39So, you know, the name of the person,
7:41the account that they're in, their their
7:43email, whatever information needs passed
7:46in order to be able to do that process,
7:48you just set it up here. And then when
7:51you go to use it u in whatever workflow,
7:54you just use a call n workflow tool and
7:57then you would choose it from your sheet
7:59here. All right. So, what do I got? Um,
8:02do I have any that I'm willing to show
8:04real quick? So, Melvin, Melvin is my
8:06research bot. All right. So, Melvin um
8:09currently is not set up, right? He's set
8:11to receive all input data. He was not
8:12set. He should not be set to do that.
8:14He's more likely to break to do that.
8:16So, just my workflow 2. How about that?
8:19So, let's save. I'm going to refresh. I
8:22set it to my workflow 2. Before I do
8:25that, actually, let's uh let's do
8:28name date ID. Now, we'll save again.
8:34refresh again. Um, and I'm going to show
8:36you what how this works. So, come over
8:38here to this guy. I'm gonna search for
8:41my workflow 2. And now I have to put in
8:45name, date, ID in order to run this
8:48process. Now, I I do not suggest
8:50creating a recursive loop like that. Uh,
8:53that will create many many issues. Don't
8:55do that. But I'm using this as an
8:58example that I could have called another
9:00workflow that asks for input. And so
9:02here, my this first bot that I showed
9:05you or the second bot that I showed you,
9:07I do that. So I have input here. All
9:08this is just processing either voice or
9:11text and pulling past conversations I've
9:13had with the bot. It then runs it
9:15through um a process. This one's posts
9:18on LinkedIn for me. This one posts on
9:19Instagram for me. This one posts on
9:21Twitter. And so I have a tool that I've
9:24built that any other bot that I make can
9:27now use that tool as well. All right. So
9:30maybe I have another bot that does
9:32research on Reddit that creates um finds
9:36the article that I want to post onto
9:38Twitter. At the end of that research
9:40bot, I can have a Twitter call to my
9:43Twitter bot, my Twitter agent that then
9:46takes the article and posts it and there
9:48it goes. I don't have I've basically
9:50fully automated that whole process. Um
9:52now I'd be a little careful automating
9:54social medias. They're checking for
9:55bots. They might ban your account, but
9:57that's just an example that I want to
9:58give you. So, I'm I'm in this video, I'm
10:01trying to really show you kind of the
10:04different ways to structure your tools
10:06because, you know, I I see too many
10:09people building one bot, just one bot
10:13that does
10:16everything. But that's not the best way
10:18to structure this. Instead, think of
10:21every process that you do for any of
10:24your workflows. You, especially if it's
10:26something that might be repeatable.
10:28Think of that as a tool that you can
10:30create in a separate workflow that can
10:32then be called by any other bot you
10:34might need that would ever use that
10:36tool. And that is really the beginning
10:39of an agentic system by having just a
10:41string of tools over and over again that
10:44you're able to run through a process.
10:45You run like and you can string them
10:48together. as an example. So I can have
10:50um um execute workflow, execute
10:54subworkflow. All right. And then after
10:57this workflow runs, I can have another
10:59execute subworkflow and just keep
11:02passing data through over through one
11:05tool after the next tool after the next
11:06tool after the next tool. And I don't
11:08even actually have to have these AI
11:09agents handling any of it. Again, it's a
11:12bunch of wasted tokens that people
11:14think, "Oh, it's so cool that I'm
11:16running it through an AI agent. It's
11:18it's not even necessarily the best way
11:20to do it." So, that's what I wanted to
11:22show you all today. Kind of a better way
11:24that I have found over the past six
11:26months of building bots um that I don't
11:29see anybody really talking about on
11:30NAND. They want to show you their nice
11:33shiny beautiful workflow that um you
11:36know honestly I you know half the
11:38workflows that I show you guys um for me
11:42they start this way. They don't start
11:44with some other trigger whether a click
11:46trigger or a schedule trigger. I start
11:48with this because it's integrated into a
11:51bigger system that I have. You know this
11:52guy here. Um I'm not only showing you a
11:55small part of him. He's got he's got
11:56tons of other things I'm building out
11:58constantly. Every day I'm building more
11:59and more parts of him and it it just it
12:02does it a text classifier. It runs
12:04through what is the user requesting? Is
12:06it any of these five things? No. Then it
12:08goes to the next text classifier. Is it
12:09any of these five things? No. Next test
12:11classifier. On and on and on until it
12:14gets the thing that I asked it to do and
12:16then it does it. That's how my personal
12:18assistant works. That's how my executive
12:20assistant works. It really identifies
12:23what I'm supposed to do and then it runs
12:24that process rather than just having one
12:27AI try to handle it all. All right, so
12:30if you found this video helpful, please
12:31make sure to subscribe to the channel,
12:32like and share the video because it
12:34really does help get that information
12:35out there to those who need it and as
12:37well as help me grow my channel. I
12:39appreciate that, too. Also, if you want
12:41additional NAD tutorials or free
12:43workflows, please feel free to join my
12:45free school community. The link is in
12:46the description below. I got over 30
12:48different workflows ranging from social
12:50media posting to executive assistants to
12:53product description writers and so much
12:56more for businesses. As always, I'm
12:58Bradford Carlton. Let's automate your
13:00success