Elizabeth Fuentes Leone and Sandhya Subramani start with a Strands Agents customer-service workflow backed by mock account and order data. A system prompt guides three tools for customer lookup, order history and refunds. Inspecting the messages and tool calls makes the reasoning loop visible before extending the example into a multi-turn conversation.
Elizabeth Fuentes Leone and Sandhya Subramani add hooks that run programmatic checks at defined points in the loop. A tool-call limiter demonstrates a hard constraint, but an audience question exposes an important distinction: blocking another tool call does not automatically end the model's turn. The desired stop behavior must be defined explicitly, rather than assumed from a prompt or partial guardrail.
Elizabeth Fuentes Leone and Sandhya Subramani introduce Markdown skills for account troubleshooting, order tracking and refunds so relevant instructions can be loaded when needed. Steering handlers add checks on workflow order and response tone. Programmatic enforcement and model-assisted guidance serve different purposes; the workshop does not establish that a buddy agent guarantees correct output.
Elizabeth Fuentes Leone and Sandhya Subramani move the customer-service code into Amazon Bedrock AgentCore. The walkthrough covers an application entry point, local testing, generated deployment infrastructure and a successful invocation. Observability requires configuration and instrumentation so sessions, traces and token use can be inspected. Workshop account offers and promotional calls to action are omitted.
Watch on YouTube
