AWS re:Invent 2025 - Amazon Bedrock Agents and AgentCore Design Patterns (TNC322)

AWS re:Invent 2025 - Amazon Bedrock Agents and AgentCore Design Patterns

Overview

• Customers are increasingly using large language models (LLMs) to build complex, integrated solutions beyond simple chatbots
• These solutions often involve integrating with external components and specialized data
• This session covers:
  • Different types of agents and agent configurations
  • Bedrock Agents and how to build agent-based systems
  • Orchestrating agent-based workflows with Bedrock Flows
  • An introduction to the new Amazon AgentCore service

Agent-Based Systems

• Agent-based systems involve multiple interconnected components:
  1. Agent workflow: Defines the actions an agent will take based on user input
  2. Components: Modular tools, models, knowledge bases, and code interpreters used by the agent
  3. Continuous evaluation: Collects metrics and feedback to improve the system
  4. Response generation: Retrieves and formats the final response to the user
• These components can be configured in different ways to create various agent-based architectures

Bedrock Agents

• Bedrock Agents provide a managed service for building agent-based solutions
• Key features:
  • Choose from pre-built agent configurations or create custom agents
  • Integrate with external tools, APIs, and knowledge bases
  • Leverage built-in capabilities like guard rails, memory management, and continuous evaluation
• Example: A banking assistant agent
  • Receives user input and checks guard rails for inappropriate content
  • Performs actions like looking up customer information, calculating loan eligibility, and generating responses
  • Uses a combination of foundation models, knowledge bases, and custom logic

Inline Agents

• Inline agents allow for dynamic configuration on each invocation
• Useful for experimentation or when the specific requirements change per request
• Developers can specify the model, instructions, knowledge bases, and actions for each agent call

Multi-Agent Collaboration

• Agent-based systems often require specialized agents to handle different tasks
• A supervisor agent can coordinate the workflow, routing requests to the appropriate specialized agents
• Example: A customer support scenario with agents for intent classification, technical troubleshooting, and policy lookup

Bedrock Flows

• Bedrock Flows provide a managed service for orchestrating agent-based workflows
• Allows defining conditional logic, data retrieval, and prompt generation
• Includes building blocks like Lambda functions, chatbots, knowledge bases, and prompt templates
• Example: A loan approval workflow that calculates eligibility, generates rejection or approval letters based on the result

Amazon AgentCore

• AgentCore is a new modular service for building custom agent-based systems
• Key components:
  • Runtime: Executes the agent logic using the chosen framework and model
  • Identity: Handles authentication and authorization for inbound and outbound calls
  • Gateway: Exposes the agent's capabilities and routes requests
  • Memory: Manages short-term and long-term context
  • Browser: Provides a headless browser for web-based actions
  • Code Interpreter: Executes custom code within the agent environment
• Allows using a wide range of open-source frameworks and models, not just those within AWS

Observability and Monitoring

• The AgentCore service provides observability features to monitor agent performance
• Developers can view active sessions, traces, and telemetry data to ensure the agents are behaving as expected

Key Takeaways

• Customers are increasingly building complex, integrated solutions using LLMs beyond simple chatbots
• Bedrock Agents and AgentCore provide managed services and modular architectures for building agent-based systems
• These solutions allow integrating external components, specialized data, and custom logic to create powerful agent-based applications
• Multi-agent collaboration and workflow orchestration enable building sophisticated, task-specific agent ecosystems
• Observability and monitoring are critical for ensuring agent-based systems are performing as intended