Name: Multi-Agent Orchestrator
Uploaded: 2024-12-20T09:15:03.040Z
Duration: 2 min 33 s
Description: A flexible and powerful open-source framework for managing multiple AI agents and handling complex conversations.

multi-agent-orchestrator

A flexible and powerful open-source framework for managing multiple AI agents and handling complex conversations.

Multi-Agent Orchestrator

He guys welcome back to the channel. My name is the here and today we'll be looking at the multi-agent orchestrator by AWS. Uh this is going to be a very short video. This is not going to be a detailed code walk through. We'll just simply be you know running the examples that are already present in this particular report that you're seeing on the screen. 

Um as you can see from the screen this looks very intuitive very easy for uh the developers to consume information. Uh this has three key components here. One is the classifier. Um essentially an intent classifier which classifies intent and selects the relevant agent. The other one obviously is an agent uh that you know processes the request using an llm and the uh the third component is memory. Right so that that also becomes a context for the classifier to essentially identify a follow-up intent or follow query and then select the relevant agent. 

Uh either the same agent that process it previously or maybe a complete new agent depending on the type of query that is coming in. All right um with that I think uh let's get started. I'm going to quickly show you some of the examples that I was running here. Let me stop this and clear screen. So I I copied the example of the fast API uh demo from the multi-agent orchestrator report. Uh and this is pretty

Simple. Uh, let's look at the, you know, the quick ones. So, there's a custom handler to essentially, uh, take in queries and then add it to a queue. And, uh, you know, you need to essentially set up the multi-agent orchestrator. There’s a lot of configuration that you can actually pass here. Uh, and the kind of classifier that you're using, there are multiple types of classifier, uh, that you can select. You can actually create your own classifier or leverage the ones that the, uh, this particular orchestrator comes with. 

Uh, you can also create agents. And if you've seen the previous demos on, let's say, crew AI or, uh, the SW open AI swamps multi-orchestrator multi-agent orchestrator framework, you will see that, uh, the agent class looks pretty similar to this. But let me quickly walk you guys through, uh, the agent class here. I'm leveraging a built-in agent called Bedrock LLM agent. You can actually create your own custom LLM agent. 

Uh, this agent takes in name, obviously the name of the agent. Do you want to really, do you want to stream the information or not? And also the description. And this is the most important one because this is what is used to identify, uh, the, uh, the intent and, you know, select this particular, you know, agent. Right? Similarly, you know, uh, I created a completely different agent called football.

Agent just to you know ask a few for queries and then there is a health agent. Let's see how this particular agent can handle the request. 

Now before I do that let me tell you the most important part of this particular demo. Uh to run the Bedrock elimination you obviously need to have access to the uh anthropic LLMs that are running on AWS. Uh to do that you need to essentially create a free free tier account on AWS and uh go to the Bedrock services and click on Bedrock configurations model access and get yourself access to these three uh models essentially Claw 3.5 HU or Claw 3.5 Sonet and in fact Claw 3 Sonet and Claw 3 HU as well. 

Uh by doing this this is free this shouldn't cost you anything and you should be able to run this on free free tier. Obviously there's a rate limit. Uh you will not be able to make several calls uh in a minute but for you to just you know try this particular agentic framework this should be good enough. 

All right um I intend to run this now let's quickly run this particular agent and see how this works. I was actually running this a few minutes ago. All right so this is running in port 8080 I have that open already here I'm going to try it out. 

So as you saw that we have created three different agents let me start with the first one. So I'm going to ask it what is AWS Bedrock.

Sorry and we'll give it a user ID and a session ID. Let's see user ID and session ID is key for you know the conversation history because you know you might want to retain the conversation history for this particular reason for this particular session. 

And then uh identify intent based on the conversation uh that you had previously using the same user ID session key to a lot of uh uh chat like interfaces right. All right uh like I said you need to set up an AWS account. 

This is referencing to thews credentials that I have set up using a CLI. I'm going to share the link on how you need to do that. Uh all right so obviously it has started running and finished processing the request. My query was what is AWS Bedrock. 

Uh it has classified the intent that and it has selected the agent to be Tech agent and the confidence score is .95 which is excellent and it has processed this in 3 seconds. So let me go back and look at the response. I'm going to increase the size here and let's see what the response is. 

Bedrock is a new service it's good. Let me give you uh you know query and see how it would identify the football agent. So I'm going to ask it who is the leading go goal scorer in international. Okay I'm not going to mention the word football and let's see actually remove the words go as well and see how.

Would it would uh essentially identify the intent. Let me go back as it as expected it has identified the intent uh to be a football agent or the intent for the football agent. The confidence score is 0.95 great and it has finished processing it in the expected time and and awesome. The Christiano obviously Ronald Christian Ronald is the leading goal scorer in uh international soccer and the information seems correct. I don't know if the G coal count is uh good. I can't validate that now but Christian is reading was correct.

So we have another agent called the health agent U. And as you can see the description is very vague. It says expert Health but let's see if it is able to identify the relevant intent. Okay remember I am using the same session ID so it Nots that I have previously asked a question on soccer or football. Let's see if it is able to you know identify the inter correctly. 

So I'm going to ask it what are the most common injuries in sports and let's let see if it will select the right intent. Yep it has selected the intent uh the relate int and selected the agent. The agent here is Health agent. The confidence score is 0.9. Uh there is a little bit of ambiguity here right because my previous question had to do with sports and then I'm following it up with uh you know query on health.

Was able to select this particular intent and the key here is let me actually first show you the response. Sprains, overuse injuries, concussions, fractures, AC tears. All right, so the key here is that the description needs to be as detailed as possible so that you know you can zero down on the precise agent or identify the relevant intent and zero down on the precise agent and execute the tasks. 

All right, you can let me show you the weather agent. I'm not going to run this. As you can imagine, every agentic framework today has the most important component tools, and this one also comes with tools. From tools, you can actually specify the kind of tool that you want to add here. 

So let me show you the tool that we've added. We've created a weather tool. There is a specific format that you want to leverage here. It takes in name, description, the input schema so that it can identify the object. One thing that I have to tell you is that the tool detail is, you know, relatively complex to set up here, at least from the example that I'm seeing versus, let's say, a crew AI. 

Because in crew AI, you just, you know, give it a doc string and then add a decorator at tool, and it still does a very good job of identifying the tool. Right, but here I think you need to create the scheme of the input and maybe the.

Output as well. I'm not sure this. Yeah, but you need to just create the scheme of input. Uh, the advantages here is that you'll have very precise input that will be generated by the uh llm. But the downside is that you'll have to, you know, do this detailed uh configuration. But you know, if you're in an Enterprise setup and you want Precision, this is a good time. Uh, this investment in time uh to write this input schema is going to be that much more beneficial in the long run.

All right, gu. Let's quickly look at the documentation as well. Um, so like I, you know, called out earlier, it has key components. The components being the classifier, uh, and agents, and the uh another component being uh, you know, the conversation storage. Um, and you can actually create custom classifiers using uh, you know, your own the llm that you have access to. That could be from Azure or, you know, Anthropic or Open AI. You can also create your own agents. There are a lot of pre-built uh agents that come in as part of the package, but you can obviously create your own custom agent. 

And the storage opportunities are also great. You can use in memory as well as Dynam Dynam dbr or, you know, set up your own custom storage. Mar gu, that's it from uh me on this one. I thought this is a very clean, very uh, you know, intuitive flow of uh, you know, data.

Uh for a developer to understand and build agentic frameworks still not as simple as let's say uh open AI swarm. But I think uh you know when you want to set up enterprise grade you know agents I think this should be an agent agentic framework that you can fall back on. 

Also open AI swarm is still not officially in production. That that's still an exper it's still in experiment stage. But uh AWS has shared this as something that you can leverage in production. 

All right guys that's it from me in today's video. If you like this video please uh you know share AI Agent Store with colleagues. Appreciate your time here thank you.

Note, that there may be slight differences in text and video. Watch video for the original content.

Multi-Agent Orchestrator Video