Build a Simple Agentic App: From Tool Calls to Safe Answers

This presentation deconstructs the repository’s Python support agent into ordinary, reusable parts. Attendees follow a ticket through the function-calling cycle—model, action, observation, next decision, and final answer—then learn to recreate that pattern using narrow read-only tools, JSON schemas, an explicit execution allow-list, and message history. The second half introduces the boundaries that make the application trustworthy: input and output validation, a six-turn limit, session isolation, safe escalation, and offline scripted-model tests. Three support scenarios demonstrate grounded resolution, incident awareness, and knowing when not to guess.

#VegasTech