Agentic AI with Amazon Bedrock and Nova

It has been over a decade since Amazon introduced Alexa alongside the Amazon Echo. Alexa took the world by storm, enabling users to interact with a “smart speaker” that could perform specific tasks based on keywords. The keywords needed to be predefined and transcribed by a service similar to Amazon Transcribe. Once parsed for specific keywords, actions could be triggered, and a response would be delivered through Echo using a text-to-speech service like Amazon Polly.

A natural evolution to this is to move to the generative AI space. The new Amazon Nova models enable the same type of interaction, but without being limited to predefined keywords. Instead, user input gains enhanced context through an MCP query, is matched against domain-specific knowledge using RAG, and is acted upon by AI agents capable of invoking specific actions based on context, research, and decision making.

As Amazon Nova continues to introduce new multimodal models that span image processing, audio, and video generation capabilities, these models provide base functionality for interactions that can be real-time and lifelike. This opens the door to a myriad of possibilities for reimagined UX interactions with automated systems.

This allows Amazon Nova to be a premier suite of foundational models capable of multimodal analysis with industry-leading price-performance and efficiency.

By combining these new models with Amazon Bedrock Flows, we can create countless new functionalities for AI Agents to automate many processes that require research, analysis, and decision-making.

In this article, we provide a brief overview of agentic AI, explain how it can be implemented using Amazon Bedrock, and walk through a sample use case of Amazon Nova models with Amazon Bedrock.

What is Agentic AI?

Agentic AI refers to AI systems composed of multiple autonomous agents that can:

• Act independently to achieve specific goals
• Make decisions and take actions without constant human oversight
• Collaborate and communicate with other agents
• Adapt their behavior based on the environment and feedback
• Maintain persistence across interactions

It is useful to connect distinct AI Agents for a few reasons. AI Agents can follow Domain-Driven Design principles, being specialized and functional in their respective domains. This means that each agent can be optimized for both domain-specific knowledge and specific tasks by utilizing different knowledge and context sources, as well as a tailored model for their task. For example, a customer interaction agent might use a text-to-speech model such as Amazon Nova Sonic, while a research agent would use another model, such as Amazon Nova Pro, to retrieve results from a RAG Knowledge Base.

Another benefit to using multiple agents is scalability and resilience. Agents can be added, removed, or altered within an agentic system without rebuilding the entire system (which is time-consuming and risky). If a particular agent is a hotspot for activity, it can be scaled up independently since the rest of the system can be parallelized. Agents can even validate the outputs of other agents, ensuring that answers are cross-checked within a system before being returned to a user or taken as an action.

The most ambitious advantage to agentic AI though, is the problem-solving abilities that this approach enables. A supervisor agent can break tasks down into smaller tasks that are assigned to individual agents who are experts in their respective domains. This scatter-gather pattern of processing means that some steps can work in parallel, accomplishing results faster. Agents can collaborate on problems and provide sophisticated reasoning across multiple “stakeholder” agents within a process.

To learn more about agentic AI, check out our blog post Getting Started with Agentic AI on AWS, which details relevant use cases and how to build agentic AI systems.

Amazon Bedrock

Amazon Bedrock is a fully managed service from AWS that provides access to foundation models (large language models and other AI models) through a unified API. It's designed to make it easy for developers to build and scale generative AI applications.

Amazon Bedrock provides some convenience features for getting started with AI:

• The Model Catalog serves as a gallery for various foundational models
• You can fine-tune or pre-train models in the Custom Models panel
• There are Prompt Routers for redirecting topical prompts to specific models
• Two playgrounds are available for tinkering with models: a text-based chat playground and an image/video playground for images
• Amazon Bedrock Guardrails provide safeguards for relevance, content filtering, and PII removal
• A number of builder tools are also available; tools for working with Amazon Bedrock Agents, Flows, and Knowledge Bases

This allows Amazon Bedrock to be a one-stop shop for all things AI on AWS, and provides one view (and one set of API calls) with all of the tools to be successful running AI on AWS, including access to multiple models. For a more in-depth look at Amazon Bedrock, take a look at Amazon Bedrock: Everything You Need To Know.

Amazon Bedrock also allows you to create agents in a few simple steps. Start by selecting a foundation model for your agent, defining the agent’s role, and providing some basic instructions along with relevant parameters. Next, configure the action groups to specify which actions the agent can perform. This can include connecting to APIs, invoking AWS Lambda Functions, or interacting with other services. You may also want to include a knowledge base for RAG or an Amazon S3 data source. Finally, enable session state storage to give the agent persistent memory for handling long-term tasks.

Amazon Bedrock facilitates memory management automatically in a few different ways. Session memory maintains conversation context within active sessions. This helps preserve user preferences and task progress, such as in a chatbot session. Amazon Bedrock also has persistent memory for storing information across multiple sessions. This remembers user history and learned preferences for long-running, multi-session projects. Amazon Bedrock also has task management memory, which persists the details of Flows. This helps track progress on complex, multi-step tasks and can resume interrupted workflows.  

With Amazon Bedrock’s automatic memory management and support for collaborative agent flows, you can chain together diverse agents to complete complex tasks. For example, building on the Alexa use case, imagine creating an interactive speech-to-speech agentic flow designed to automate call center tasks. To create a conversational AI, you would first start with a Supervisor Agent, which is a central AI entity that manages and coordinates (orchestrates) multiple AI agents to achieve a complex task or workflow. The supervisor agent could use the following agents and knowledge bases:

• Speech-to-text agent - Processes voice inputs into text for the other agent.
• Knowledge agent with RAG - Uses retrieval-augmented generation and a knowledge base to search relevant documentation or policies based on the user’s inquiry.
• Customer experience agent - Checks results against a customer profile to ensure to ensure relevance, enforces data access using security roles, and applies Amazon Bedrock Guardrails for content filtering.
• Response generation agent - Crafts a personalized solution tailored to the user’s issue.
• A text-to-speech agent - Translates the response into voice in real-time. 
• Customer insights agent - Each interaction is logged to the customer’s profile, providing future agents with a history of past conversations and additional context about the customer account.

This multi-step process is fully supported by Amazon Bedrock and can deliver a significantly better experience than traditional IVRs and call center routing systems commonly found in the market. Unlike a call center tree or Amazon Alexa, AI can decide the path to take for the user without a predetermined if-this-then-that type of logic attached.

Amazon Nova and Agentic AI

Circling back to the Amazon Nova, the above scenario is possible using the new models. From the Amazon Documentation, the following models are available:

• Amazon Nova Micro / Lite / Pro / Premier for text, knowledge, image, and video inputs.
• Amazon Nova Premier for decision making and knowledge distribution
• Amazon Nova Canvas for image generation, editing, and adjustments
• Amazon Nova Reel for generative video
• Amazon Nova Sonic for text-to-speech, and speech-to-text.

This means that if we were mapping the previous flow to Amazon Nova models, we would choose:

• Amazon Nova Sonic for the speech-to-text agent
• Amazon Nova Pro with RAG for the knowledge agent with RAG
• Amazon Nova Pro for the customer experience agent
• Amazon Nova Pro for response generation agent 
• Amazon Nova Sonic for a text-to-speech agent
• Amazon Nova Premier for customer insights agent

For the agents using Amazon Nova Pro, it may be possible to reduce costs by opting for Amazon Nova Micro or Amazon Nova Lite, if the context window or accuracy can be smaller than Amazon Nova Pro.

This means that you can keep the entire multimodal call center application inside of the Amazon Nova Ecosystem, which can offer substantial cost efficiencies at scale. On average, Amazon Nova models offer a 70% cost savings per token compared to GPT or Claude models. Let’s say that each customer interaction uses roughly 10,000 output tokens on average, and the call center application handles 5,000 calls per day.