Strands Agents tutorial: build and deploy to AgentCore

Elizabeth Fuentes Leone and Sandhya Subramani build a customer service AI agent with Strands Agents, then deploy it to Amazon Bedrock AgentCore. They describe Strands as an open source harness that supports different model providers and local hosting. The workshop uses a Bedrock-hosted model; it does not demonstrate an offline setup.

The example starts with mock records for two customers and tools for customer lookup, order history and refunds. A system prompt defines how the agent should use them. The presenters inspect messages, tool calls and token metrics, then use the same underlying code for a multi-turn chat.

A hook limits tool invocations per request, initially to three. Lowering the limit exposes an important distinction: blocking a tool call does not end the agent's turn. The demonstration continues its model loop, and the presenters explain that ending the turn requires a different hook behavior. Markdown skills supply procedures when needed, while steering handlers check the refund workflow and response tone.

The deployment section covers an application entry point, local testing and generated AWS infrastructure. The presenter says local use of AgentCore and the demonstrated Bedrock model requires AWS credentials. For observability, she enables transaction search and adds an instrumentation dependency, then shows a successful invocation and where to inspect sessions and traces.