A travel company wants a conversational agent that answers policy questions from its help center, checks live booking APIs, and books changes when the user confirms. Responses must cite sources and invoke verified backend actions. Which AI agent architecture integrates tools using Bedrock AgentCore and Knowledge Bases?

The correct design succeeds because a Bedrock agent orchestrates reasoning over conversation history, calls a Knowledge Bases retriever to ground policy answers in help-center chunks with returned citations, and exposes live booking lookup and modification as Action Groups with OpenAPI schemas backed by Lambda, which enforces parameters, authentication, and a confirmation step before state-changing calls. A RetrieveAndGenerate workflow with Lambda resolvers is a legitimate RAG pattern for Q&A plus lightweight integration, but it lacks the agent loop for planning, tool selection, and confirmation handling, so it cannot reliably sequence check-then-book with verified actions. An agent with Action Groups plus direct S3 prompts is also common for small corpora, yet pasting documents consumes context, loses managed chunking, vector freshness, and automatic citations, failing the sourced-answer requirement. An agent with retrieval and Action Groups that books immediately on detected intent looks fully integrated but violates human-in-the-loop governance for travel changes by skipping explicit confirmation, creating risk of unintended charges. For the exam, remember that Knowledge Bases alone implies retrieval while Action Groups imply schema-governed function calls, and only the agent combines both with orchestration. As a practical check, inspect the agent trace in the Bedrock console to verify a retriever citation precedes the answer and that the booking Action Group returns a confirmation prompt before Lambda executes with least-privilege IAM.