AWS re:Invent 2025 - Build multi-modal AI agents with Strands Agents and Amazon S3 Vectors (DEV332)

Building Multi-Modal AI Agents with Strands and Amazon S3 Vectors

Overview

This presentation covers a new AI agent framework called Strands, which aims to simplify the development and deployment of production-ready AI agents. The speakers, Joy Chakraotti and Elizabeth Leoni, demonstrate how to build multi-modal AI agents using Strands and integrate them with other AWS services like Amazon S3 Vectors for reliable, long-term memory.

Key Benefits of Strands Agents

• Simplifies Agent Development: Strands reduces the amount of "plumbing code" required to build even simple agents like chatbots.
• Modular and Testable: Strands follows a model-driven design approach, making the agent code more modular and easier to unit test.
• Evolves with LLM Advancements: The model-driven design allows the language model to determine the best execution strategies, insulating the developer from rapid changes in the LLM landscape.

Strands Agent Architecture

• Core Primitives: Strands agents are built on three core primitives - Prompt, Tool, and Model.
• Agentic Loop: The agent orchestrates the execution of a task by iterating between the model (which analyzes the prompt and decides on a strategy) and the tools (which perform specific functions).
• Tool Integration: Strands supports both local Python functions and remote MCP (Model Context Protocol) tools, allowing agents to access external services and data.

Strands Agent Examples

1. Simple Chatbot: Creating a basic chatbot agent by instantiating the Agent class and using default configurations.
2. Customizing the Agent: Demonstrating how to set a preferred model provider, model ID, and system prompt to customize the agent's behavior.
3. Specialized Agents: Building an AWS solutions architect agent by setting a specific system prompt.
4. Multi-Modal Agents: Integrating image, file, and video analysis tools to create a multi-modal agent capable of processing diverse content.

MCP (Model Context Protocol) Tools

• MCP Server: Exposes local Python functions as remote tools that can be accessed by the agent using the MCP protocol.
• MCP Client: Allows the agent to communicate with the remote MCP server and utilize the exposed tools.
• Benefits: MCP tools enable agents to leverage external services and data sources without requiring custom integrations.

Conversation Management

• Agent State: Maintains stateful information that persists across multiple agent invocations.
• Request State: Stores contextual information within a single agent request.
• Conversation History: Tracks the full conversation history between the user and the agent, enabling the agent to remember and reference past interactions.
• Conversation Manager: Provides options to save and manage conversation history, including session-level and user-level persistence.

Amazon S3 Vectors for Long-Term Memory

• S3 Vector Store: A new AWS service that allows storing and querying vector embeddings in Amazon S3 buckets.
• Integration with Strands: The presentation demonstrates a custom Strands tool that leverages S3 Vectors to store and retrieve conversation history, enabling long-term memory for the agent.
• Benefits: S3 Vectors provide a cost-effective and scalable solution for maintaining persistent memory for AI agents, complementing the short-term conversation management capabilities of Strands.

Key Takeaways

• Strands simplifies the development and deployment of production-ready AI agents by providing a model-driven, modular, and testable framework.
• Strands supports both local and remote tool integration using the MCP protocol, enabling agents to access external services and data sources.
• Conversation management capabilities, including state, history, and session persistence, help maintain context and memory across agent interactions.
• Amazon S3 Vectors offer a scalable and cost-effective solution for long-term memory storage, complementing the short-term conversation management of Strands agents.
• The combination of Strands and S3 Vectors enables the creation of sophisticated, multi-modal AI agents with reliable, persistent memory - a key requirement for real-world, production-ready applications.