In this talk I will explore the reality of what happens when you try to build an AI agent on AWS using Amazon Bedrock and tool calling, inspired by still new and emerging AgentCore patterns. Through a live demo and practical examples, I'll walk through what worked, what didn't work, and the unexpected challenges that appear when moving from theory to implementation, including hallucinations, incorrect tool usage, and trade-offs around cost and latency.

The goal is to give a realistic, engineering-first perspective on building agentic systems today, and help developers understand where these patterns are useful and where they still fall short.

By the end of this talk, you'll have a clear understanding of how to approach building AI agents on AWS, what pitfalls to expect, and how to make informed decisions when applying agentic patterns in real-world systems.

What you'll learn

• Identify the specific failure modes - hallucinations, incorrect tool selection, cost spikes - that appear when Bedrock agents move from demo to production.
• Apply practical patterns for tool calling that reduce error rates and improve reliability under real traffic.
• Evaluate cost and latency trade-offs when designing agentic systems so you can make informed architecture decisions.
• Debug unexpected agent behaviour using the observability tooling available in Amazon Bedrock today.

About the Speaker

Marina is a Tech Lead at Vidatec working with web, mobile, MS Teams and AI solutions across multiple industries. She has significant experience working with AWS services, managing deployments and maintenance across multiple production workloads. Marina is a founder of Catbytes - online community for women in tech and a co-organiser of AWS UG UK and AWS AI In Practice community events.