Amazon Bedrock: from RAG to Agents

You will build generative applications on AWS, starting with Amazon Bedrock fundamentals and finishing with a real conversational agent on AgentCore Harness — without writing an orchestration framework from scratch. You will learn how to move from Bedrock Agents, now in maintenance mode, to AgentCore through a practical case: a restaurant-booking concierge.

• Developers with intermediate Python and no prior experience of Bedrock or AI agents
• People who have already used a model through an API, such as OpenAI or Bedrock, and want to move from a single LLM call to an agent with memory, tools, and RAG
• Backend and cloud engineers who want to understand the difference between calling an LLM and orchestrating a managed agent
• This is not a machine-learning or model fine-tuning course. The focus is integration and architecture, not training

Module 1 — Bedrock fundamentals and prompt engineering (pre-work, before the session). Set up the boto3 client, survey the serverless model catalogue (Amazon Nova, Claude, and open-weight models), and practise basic prompt engineering. This module provides the foundations needed for the live session.

Module 2 — Knowledge Bases and RAG. Build a Bedrock Knowledge Base with vector retrieval and compare OpenSearch Serverless with the lower-cost S3 Vectors option. This prepares you to build agents that need specific domain knowledge.

Module 3 — LangChain and ReAct agents. Use LangChain (LCEL) to orchestrate prompt chains and build agents with tools through create_agent, LangChain’s LangGraph-based API. This provides the conceptual bridge between a single LLM call and a complete agent.

Module 4 — Your first agent with AgentCore Harness. Learn why classic Bedrock Agents is now in maintenance mode and build an equivalent agent declaratively, defining its model, tools, and memory without writing an orchestration loop. The practical case is a restaurant-booking concierge.

Module 5 — Multiple agents. Use the agent-as-tool pattern to make a generalist agent collaborate with a specialist, without a custom orchestration framework.

Module 6 — Advanced AgentCore. Add persistent memory, reusable skills, and a code interpreter. Test a governance policy in practice by applying a guardrail to the concierge, then review how automated evaluation and observability support a production-ready agent.

The modules progress from the 0/1 foundations in the pre-work to the Module 6 capstone. Modules 4, 5, and 6 form the practical core of the in-person session.

• Configure a managed agent on Amazon Bedrock AgentCore Harness (model, tools, memory) without writing orchestration code
• Build a RAG Knowledge Base and choose between OpenSearch Serverless and S3 Vectors based on cost and scale
• Connect an agent to external data and functions through tools (inline and via Gateway)
• Add persistent memory and a policy guardrail to an existing agent, and know where to look for observability and evaluation in production
• Have a generalist agent collaborate with a specialist agent through the agent-as-tool pattern

• Intermediate Python: you should be able to read and write functions and be familiar with pip and virtual environments
• A working AWS account with access to the Bedrock console. Setup and model-enablement instructions are sent before the workshop
• No prior experience with AI agents or Amazon Bedrock is required. Modules 1 and 2 are designed for people starting from zero and must be completed as pre-work before the live session

• Pre-work is about 1 hour (Modules 1 and 2). The live full day covers Modules 3, 4, 5, and 6
• Exercises run on notebooks. Bring a laptop you can use to work in an AWS account