amazon-bedrock-agents

A fully managed service that enables developers to create and manage AI agents capable of automating complex tasks by integrating foundation models with APIs, data sources, and applications.

Amazon Bedrock Agents

Welcome. My name is Mark Roy and I'm a principal ML specialist solution architect at AWS. In this brief video, I'm going to take you through an overview of agents for Amazon Bedrock as well as give you a quick demo of how it works. So let's jump right in. 

Uh, first of all, let me cover what is agents for Amazon Bedrock. Uh, and as you probably know, uh, generative AI has been being used by pretty much every customer and across all use cases and industries. So it's, uh, and it's been shown to be extremely powerful. One of the things it's not doing, however, is taking action. So customers are looking to build solutions using large language models services like Amazon Bedrock, but they want to take it to the next step. 

Uh, so what agents allows you to do is build these applications that can execute multi-step business tasks. You can interact with it in natural language, take advantage of the full power of LLMs, large language models, uh, under the hood, but also take actions by taking advantage of your existing APIs and applications, databases, and, uh, knowledge stores. 

So a quick look rundown of the features, uh, that agents provides. One, it uses the power of Bedrock's large language models. Uh, secondly, and this is really where the secret sauce comes in, uh, it's able to take your request, break that down into a set of steps, and execute that on your behalf. So it's

Not a hard-coded application. Uh, it's actually coming up with a plan on its own given a set of tools or actions and knowledge bases and then executing that for you. Uh, it's also done in a fully secure fashion. This is a fully managed service and you get full control over for who can use which uh agents and actions and knowledge bases. 

And then lastly, uh because this is a a a uh it's making decisions uh under the hood about how to execute it also gives you access to uh what that plan was and how it executed that uh transparently. So because the orchestration is something kind of new uh and interesting let me dive into that a little bit further. 

So when you use an agent you hand it a task and agents for Amazon Bedrock decomposes that. It figures out a plan what are the steps that I need to do to get that job done on your behalf. Uh, and based on that plan it will execute uh various actions or do some uh searches across uh multiple knowledge bases. It will look at the results that it gets and proceed to the next step until it's got uh uh the U task complete until it has a final answer. 

Let's double click on that a little bit more uh deeply and then I'll jump into a demo. So here we've got a Bedrock agent that's been handed a task. Uh, and under the hood it does all of the heavy lifting of constructing a proper LLM prompt. It takes into account.

All of the conversation history you've been having with the agent. It gives the metadata about which action, what tools are available, and what knowledge bases are available. It gives it the set of instructions that the agent builder provided. Uh, and then it will invoke uh Bedrock under the hood. And then uh, the agent will figure out a plan using what's called Chain of Thought uh reasoning here. It will execute each of those steps using either an action or a knowledge base. When it gets the full set of results, it will then use Bedrock again uh to prepare a final response using natural language uh back to the actual user. Very powerful new way to go about taking full advantage of uh large language models.

So let's see a demo of how that works. Uh, so what you're looking at here is the Amazon Bedrock console. I've clicked on agents and I've brought up uh an agent that I've already built. This agent is a CRM agent or customer relationship management agent. Uh, that makes it easy to uh leverage um the data that's available uh in your CRM. So the uh, let's say I'm an account manager and I work with multiple different customers. Uh, the CRM lets me keep track of who those customers are, the opportunities uh that I'm working on uh with them, as well as the history of all the different uh activities. Let's say meetings and discussions and events.

And so forth. Uh so all of that's available. So uh when you define an agent, you give it a name. Uh and uh here I've got a set of versions that I've been building. 

Let's dive into the current working draft. So you've got the name of your agent and then a set of instructions. Uh so here I'm saying that you are a customer relationship management agent that helps salespeople plan their work. Uh so forth and so on. Uh and you also identify which Amazon Bedrock large language model you're using. In this case, we're using CLA 2.1. 

And then you define a set of knowledge bases you want to make available to it, as well as uh a set of actions or APIs that are available. In this case, I'm just giving it a couple of actions. Uh one, an ability to get back a company overview. Uh a second one that lets uh the agent get back the set of recent interactions or activities. 

Uh and give details back about that, like for example, um uh an array of um dates, uh times, uh as well as um uh the meeting notes uh from those interactions. And then lastly, uh it has an ability to get the customer preferences. So these are the uh actions that are uh available to this agent. 

And let's see how the agent uh works. So you can either use the agent from the test console as you're developing it or you can use the API. So here's an example call that I made uh before we started.

Asked the agent to help me out with uh customer CABC. Given our recent interactions, what are their top concerns and interests? It came back with an answer saying hey they're worried about reliability and they have a strong interest in our upcoming service uh particularly around pricing. 

If I show the detailed trace of what went on uh behind the scenes, you can see that it came up with a multi-step plan four steps and then it executed uh those steps. The first step was to find the interactions. It passed in the customer ID which it was able to uh pull from my request. 

And then it uh invoked uh the action which interacted with the CRM itself. It gave back a set of details here about all the meetings that we've had, when they happened, what the results were. Uh and then it uh determines that it's got enough information and it gives back a final response basically summarizing what were the concerns based on uh that activity. 

Uh so pretty powerful and then you can actually have a conversation as well. So here I'm going to have a follow-up request based on those interests. Suggest a next meeting topic and an agenda and let me know what day and time is preferred. 

So it's able to leverage what it has already seen to come up with a meeting topic which is basically uh an overview of the smartship capabilities and it proposed an agenda uh smartship.

Overview pricing and packages demo Q&A. Uh you notice it came back with a uh suggested meeting time of Wednesday afternoon and that was based on its ability to uh take an action on my behalf to look up the customer preferences. 

So I had all of this powerful interaction with really just a couple of simple natural language requests and agents for Amazon Bedrock uh is able to interpret those uh and take action on my behalf. Dynamic Ally coming up with a plan based on what are the actions available and what are the knowledge bases uh available. 

Taking that a little bit further in addition to using it from the test console as you're building an agent you can also uh write your own uh application or integrate an agent with your own application. Uh you could use it from a script uh you could use it from an existing um tool or website or your uh you know uh iOS application and there's a full API to invoke the agent uh as well as create agents uh as well. 

So in this case um I built a simple user interface that takes in a request it calls invoke agent uh because we deployed that agent and made it available uh to the API. Uh and as it it's uh proceeding the API gives you back a trace of what's happening. 

So here you see the steps that it was planning to do uh and then the first step that it uh took on my behalf what parameters it passed what.

Results it got back and then it takes another step. Um, comes back with some results and then it figures out it's got enough information. So it, uh, returns that back, uh, to the invoke agent, uh, response. 

And my application would really just show this. Uh, I could save this trace information into a log file if, uh, if I wanted to, for example. Uh, so you can set up a bot of your own, uh, that will take natural language input, use one or more agents under the hood, and the agents will, uh, take advantage of your existing APIs and databases and document stores, uh, to get work done. 

Very simple to build. Um, in the way that you hook into your databases and APIs is to write simple Lambda functions, uh, in whatever language, uh, you'd like. In this case, I've got a Python Lambda function. 

Uh, so we take in an event. The agent hands you an event. Uh, you can look at what action the agent wanted you to take. In this example, uh, if I am asked to list recent interactions, call a function, uh, that will get that done. 

In this simple demo, I've just, uh, retrieved the customer ID parameter and the comp parameter, and then I've got a hardcoded set of, uh, meeting results that I'm handing back. In the real world, this is where I would call the CRM API, uh, to get that job done. 

And, uh, then I simply, uh, return, uh, that back as part of the action response, and the agent, uh, API takes care of the.

Rest so very straightforward to use very powerful new way to look at building applications using large language models. So with that, uh um let me also point out that we've got some resources for you to help you get started. 

Uh on the left, uh this will get you to our uh announcement blog post, uh to help you understand uh a little bit more about agents. And on the right, we've got a an online workshop, uh that you can uh learn about Amazon Bedrock overall with a set of different modules including a module specifically on agents for Amazon Bedrock. 

With that, thank you for watching and uh let us know what you go out and build with agents for Amazon Bedrock.

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

Amazon Bedrock Agents Video