A lot of AI agent demos feel like medieval pageantry: impressive banners, dramatic promises, and one weak gate away from collapse.

This talk is about what happens after the demo. For engineers, architects, and tech leads exploring agentic systems, we’ll turn a flashy but brittle prototype into something you could actually trust inside a real organization. Along the way, we’ll use a simple mental model for AI agents: not magical workers, but software systems that need boundaries, state, permissions, feedback loops, and fallback paths.

The session includes a live demo that starts with a fragile version of the system and progressively hardens it with guardrails, observability, retries, and human-in-the-loop checkpoints. Attendees will leave able to explain when an agent is the right pattern, compare agents with simpler workflow designs, implement safer tool-use boundaries, evaluate reliability with meaningful checks, and troubleshoot common failure modes before they become production incidents.