1. Types of AI Agents🤖 - SCAI017

All right, so let's talk about the actual agent itself. So when we come in here, let's add a chat message trigger real quick so we can add an agent to it. And we have advanced AI, got different options here. AI agent, open AI, you've got different chains you can use, text classifier, all this kind of stuff. So we're gonna grab an AI agent, of course. And so from here, first thing to note is that we have different options as far as the actual agent we wanna use. There's six options, tools, conversational, open AI, plan and execute, react, and SQL. And so full transparency here, I pretty much only ever use tools and conversational. A few times I've been able to get my workflows to actually work better with a plan and execute compared to a tools. But we're gonna go over them at a high level. They're all a little bit different with the way that they work and sort of their prompting and their options. So we'll cover that. But before I dive into that, I just wanted to throw out there. Tools, always what I go with first. If maybe I'm making some sort of chatbot or interfacing agent, I might go with conversational, depending on how many tools are hooked up. But tools works most of the time. And then sometimes if tools isn't working right and I wanna maybe switch it up, I'll try a plan and execute or potentially react, but mainly plan and execute. I haven't done too much with reacts. But this just goes back to that aspect of when you're creating a project or when you're delivering on a project, at the end, you've got sort of like you're 80% there and then bridging that gap between 80% done and then 100% production. That last 20% is really tough when it comes to so much back and forth. Testing out tons of different inputs, testing out different models, of course, like large language models. And then we're testing out different types of agents. And then we're testing out different prompting based on what we get back from the agents and the way that we see their logs. So testing out, switching up the different agents at times is helpful. But one thing to note, if you have a system message, if you have a system message and then you switch the agent, the system message goes away and you lose it. So copy that before you switch. Anyways, tools agent. So the purpose is a general purpose agent that has various tools and you're gonna define what those are and it's instructions of when to use each one, what the tools do, how it can ultimately achieve the end goal that you're looking for here. Like I said, most of the time, I'm gonna start off with a tools agent and only switch if I notice that I need to. Best when you have the agent has access to different APIs, different tools, different agents that you built out as workflows, different workflows that you built out as tools, that kind of stuff. And yeah, as far as getting them to work right, it really just comes down to effective prompting here, obviously within the system message. So one thing to understand is the difference between a system message and a user message. So I'm not, I think this may be newish that they threw this in here, but it's, you know, so the system message is where you actually prompt the agent as far as, you know, your helpful assistant. This is your, the context of the situation you're working in. This is your goal. These are the tools you have. This is when you use each tool and then here are some examples of, you know, user messages you may get and then outputs or the steps that you should take then and then some outputs like that. That system message is kind of, you know, more defined. It's like the persona of the agent and then the user messages are what comes in every interaction every time you chat to the agent. You know, if I said, let me just quickly connect a model. Hello. Okay, yeah, we're good. If I said hello, we can see in here that's the user message coming in is hello and helpful assistant is still your helpful assistant. So that's a quick difference. Let's move on to conversational. This one is gonna be used for more handling like human-like natural conversation. That's why it's called conversational. You maybe, you could optimize this for chat style interactions because, you know, it's obviously gonna remember context and not that other agents can't remember context but this one's gonna be more conversational. So if you're building like some general conversational things like chatbots or customer support, answering FAQs, conversational user interfaces, definitely consider testing out a conversational agent but keep in mind, maybe the more tools you add on you could wanna try to go back to a tools or a plan and execute. One thing to note here is that we have a different option. We have a human message. We have a human message and also we lost an option which was like pass through binary, pass binary images. So that's one of the changes there. So we have a human message now as well as a system message. So let's look into here real quick. The human message says, now here are the tools that the assistant can use. The assistant can use these tools to look up information to answer the user's original questions. So this could be, you know, like your database of FAQs, product information, give it some instructions here. And then it's interesting because we're also gonna be feeding in the user input to this conversational agent, you know, even though it's kind of getting in here but I guess now we're defining what it is. So, yeah, so that's an interesting aspect but then of course we have our system message which is more of the, you know, normal system message, what the assistant does, what it's designed to do, current and constant learning and improving. It's able to generate its own text based on the input it receives. Overall, powerful system that can help with a wide range of tasks. So that's sort of the system message here. Yeah, so conversational agent could be better test out with human-like conversational aspects. OpenAI functions agent. So I've only played around with this one a little bit. It's got the same options as the tools agent except for not the binary images part but system message and max iterations, return immediate steps. So from here, when you would wanna use an OpenAI functions agent, if you wanna utilize OpenAI's ability to execute predefined, those predefined functions, those dynamic OpenAI functions because obviously this agent's gonna take advantage of those function calling to trigger specific operations, obviously, of course, based on the input that you're getting from user. Okay, now going on to a plan and execute. So this is an interesting one because, you know, this is kind of designed for more high-level tasks that have multi-steps of reasoning and planning and then executing. Executing, executing. Capabilities of this agent is that it's going to sort of break down the main query coming in. It's gonna understand the tasks, break them into smaller subtasks, and then sort of go in a logical order. So with the video that I just made where I did, like, I built like a research team to do newsletter automation, this could have been a good use case to use a plan and execute agent. I just ended up wanting to do a sequential because it made more sense because it happens in the same order every single time, but it had a logical flow. It had a logical order as far as, okay, once the form comes in, we got a newsletter expert that creates the table of contents based on initial research, hits it off to the project planner. Project planner hits it off to the research team. Research team sends it to the editor. So that's sort of like a logical linear flow with something like more of a personal assistant. Hey, can you grab my events for today? Send an email to Mark and then do this. Sure, maybe the logical flow is the order of the query, but sometimes it's different. I don't know if you're looking up contact information, but it's just much more variable as far as the tools that it has rather than kind of every time you're gonna do this flow. So I don't know if this would be hooked up to something like the trigger here could be every time a row is updated in CRM or a new lead comes into the CRM. And then there's gonna be sort of like, okay, well, we get a lead. What do we do with it? We do this and then we push it into here and then we send an email and then that kind of stuff. So I say that because you don't actually give it really like a prompt. So here's all that we have. The only option is a human message template. As you can see, I clicked add option and this is the only one that popped up. So we've got previous steps and then you have the previous steps. And this isn't anything that we really touch. I mean, you can, but whenever I've tried to sort of fill out some information and give more detail here, it just doesn't work as well. So I would pretty much just like kind of add option and then leave it when I was testing out with these agents. But anyways, this is kind of like, as you can see here, agent scratch pad. So this is almost more for like the agent to be able to reference as it's going through the process. It's like its own sort of like internal log where, hey, this is what I'm currently doing because of this is my current objective. This is what I just did. And then here's like what I'm gonna do next based on the tools and their descriptions and the input that I got. So yeah, that's pretty much how this one works. Like I said, I usually don't start here, but I start with the tools. And then if I am getting frustrated, maybe I'll try a plan and execute. And then these last two, I haven't used too much, but React, the purpose is like, they say it's supposed to be sort of for like a reasoning framework where the agent reacts to the inputs and decides the next steps autonomously, which I kind of also think the tools agent is the same thing. And as you can see, it's got a few different, what's it called? A few different ways that you prompted in here. And that's why I don't use it because we don't have like a general system message. We've got the human message template. We've got a prefix message, which it says answer the following questions as best as you can. You have access to the following tools. You get a suffix for the chat model, and then you also have a suffix message for the regular model. And then finally, you know, return the eight steps. So this is why I haven't played with it too much. It just seems like a lot to put in here when a tools agent can get the job done with just like kind of one refined system prompt. Maybe these are a lot more powerful if you really understand like how exactly to put in these different suffix messages and prefix messages. So something I should play around with, but so far, I mean, everything has been working with tools and plan and execute. So that's the React. And then we obviously have SQL, which is, you know, kind of obviously designed to interact with SQL databases. As you can see, you connect a different source, Postgres, MySQL. Haven't done much for SQL Lite. Connect your account. And then, you know, you obviously have to define your input once again. And then you have different ways you can, you know, add some options in here as far as like schemas that you're gonna be looking at, what to pull back, what to not. Once again, we have a prefix prompt and a suffix prompt, and then we have a limit. So in here, we can see what the prefix prompt is. You are an agent designed to interact with the SQL database with, okay, I guess maybe they want you to pronounce it an SQL because it is N, but I typically just say a, or not a, I just typically say SQL rather than SQL, a SQL, okay, anyways. Given an input question, creatively, or create a syntactically correct query to run and look at the results of the query and return the answer. Do not make up any statements. If the question does not seem related to the database, just return, I don't know. So, you know, you've got this, right, but you could also just kind of create a natural language to SQL processing agent. I'm not sure if that's better or worse than this. I think that the, maybe there's just some sort of base prompts already put in here about, hey, like you are always kind of like a SQL. I mean, it's nice, you've got your data source really nicely integrated in here and some other predefined options. So that's sort of the use case of when you do that. So those are the six main types of agents that are in here. Like I said, 90% of the ones I do are gonna be tools, play around with conversational, and then if you maybe are not getting the tools the way you want it to, and you can see, it's like a nice logical flow of what's going on with your tool call based on the input, maybe try out plan and execute, but just keep in mind that they're all kind of different and that they have different ways that you configure them. So yes, it's nice to understand high level what they each do, but ultimately based on the use case, you may have to just run it five times as a tools, run it five times as a conversational and just see what outputs you like better and see the way that it took place in your execution logs. So that's gonna be it for this one. We'll keep diving into more about the actual agent configuration and stuff like that, so.