© Carbon60 Operating Co LTD
© Carbon60 Operating Co LTD
In a previous blog, Demystifying AI Agents: 6 Types and Characteristics, we explored what Artificial Intelligence (AI) agents are, the different types that exist, and their distinct characteristics and capabilities. We covered six different types and their characteristics:
Simple reflex agents: These agents operate on predefined rules, making them straightforward to implement.
Model-based agents: This model, often powered by a Large Language Model (LLM), allows the agent to analyze the context of a situation and make more informed decisions.
Goal-based agents: These agents utilize reasoning and planning to determine the best course of action to reach their goals and solve multi-step problems.
Utility-based agents: Going beyond simply achieving goals, utility-based agents aim to maximize a specific measure of "goodness" or utility.
Learning agents: Learning agents represent a significant step towards autonomous AI and can evolve and adapt through experience.
Hierarchical agents: Going one step further with individual learning agents, hierarchical agents introduce a level of sophistication by organizing these agents into coordinated teams based on a hierarchy.
Below, we’ll dive into practical implementation, focusing on how to bring these agent types to life using Amazon Bedrock.
Amazon Bedrock is AWS's dedicated portal for building and deploying generative AI applications. It offers a streamlined ecosystem, starting with serverless model hosting of a diverse range of foundation models. This includes popular third-party models from leading AI providers such as Anthropic and Meta's Llama family, as well as users' ability to import and host their custom models.
Unlike Amazon Q, which focuses on providing conversational AI for specific use cases such as information retrieval and code development, Bedrock empowers developers to build a broader spectrum of generative AI applications.
This growing ecosystem also features a suite of tools for LLM-powered workload development, including intuitive experimentation playgrounds, robust building tools, essential safeguards, and optimized inference capabilities.
While the Bedrock IDE—now integrated into SageMaker Unified Studio—enhances the overall development experience, our focus today is on Bedrock's core building capabilities. In particular, we’ll explore the creation of powerful Bedrock agents.
At OpsGuru, we believe in a use-case driven approach to development. While emerging technology draws attention, its true value lies in its ability to drive tangible outcomes. We'll demonstrate this through a practical example as we explore Bedrock Agent.
AI agents excel at delivering personalized experiences, scalability, and 24/7 availability, making them well-equipped to aid in gold trading efforts.
Gold traders rely on many factors to inform their decisions: economic indicators like interest rates and growth, geopolitical events, industry news, market sentiment, and historical price trends. A gold-trading analyst can aggregate and analyze this information, empowering traders to make informed decisions.
In this simplified workflow, our initial design uses a GenAI Analyst agent to collect and process relevant data, while leaving decision-making authority to human traders. However, depending on risk tolerance and desired automation levels, this foundation can be readily extended to greater automation.
For the GenAI Analyst, the workflow can be further broken down into the following process:
The user sends a request to the agent.
The GenAI Analyst analyses the user request.
The GenAI Analyst orchestrates tools to get the information needed to answer the question.
Simultaneously, an API call is made to check the latest price, the historical gold price, and the Gold Industry News Archive.
The GenAI Analyst analyses based on the information retrieved in step four.
The GenAI Analyst constructs output.
The user receives the output.
This use case aligns with the learning agent design discussed in the previous blog, where a knowledge base, Retrieval Augmented Generation (RAG), feedback mechanisms, and memory augment the agent. Our gold-trading assistant includes the following key components:
Real-time Gold Price (RAG): We'll use an agent action group on Bedrock to make API calls. This real-time data retrieval exemplifies Retrieval Augmented Generation (RAG).
Historical Gold Prices (Knowledge Base): The most cost-effective approach is to store historical gold price data in a knowledge base. We'll utilize Redshift Serverless, a data warehouse that integrates seamlessly with Bedrock agents, to store and retrieve this structured data.
Gold Industry News Archive (Knowledge Base): A text-based archive of gold industry news is another crucial knowledge base component. We'll use OpenSearch Serverless, a vector database well-supported by Bedrock agents, to store and access this information.
Bedrock agents offer built-in memory management. By default, session summaries are retained for 30 days and can influence future interactions. This duration is configurable. Additionally, real-time user feedback further enhances the agent's utility.
Naturally, an LLM is essential to support this workflow, facilitating text analysis and generating responses. However, the agent also needs clear instructions to leverage the right tools and fulfill its purpose. AWS Bedrock Agent provides default templates for:
Preprocessing: Parsing user input to determine if and to whom information should be passed. This helps avoid malicious intent and wasteful orchestration.
Orchestration: Defining how actions should be executed, including calls to Lambda functions and knowledge bases.
Knowledge Base Response Generation: Controlling how and if knowledge bases (such as Opensearch and Redshift) are included in responses.
Post-processing: Formatting and presenting the final response to the user.
This is a sample screenshot of what the setup in this walkthrough looks like.
We've included a screen recording showcasing a real interaction to showcase the agent's capabilities. As you'll see, the current price of gold is dynamically fetched via the Lambda function, while the footnotes reference documents are retrieved from the knowledge bases.
This demonstration highlights the power and flexibility of Bedrock Agent in building a sophisticated assistant. We've created an AI-powered tool that can empower traders with real-time insights and historical context by combining LLMs, RAG, knowledge bases, and agent instructions.
The short answer is no—this standalone agent is just a starting point. It can be integrated into more complex workflows as part of a hierarchical multi-agent system. We'll explore these possibilities in future posts, examining services like AWS Bedrock Flow.
Implementing AI agents on Bedrock opens a gateway to building truly intelligent systems. As we've seen, the power of hierarchical architectures, combined with the reasoning capabilities of modern AI models, allows for nuanced planning and execution.
This trend reflects the broader shift in AI towards systems that can think, plan, and adapt. By embracing these advancements, developers can create AI agents that perform tasks, reason and strategize, pushing the boundaries of what's possible. The future of AI lies in these sophisticated, multi-layered systems, and Bedrock provides a strong platform to build them.
today to get in touch with our team of experts to find out how Bedrock Agent can improve your business process.
