The Demo Worked. Then Real Life Happened.

Getting an AI agent to work in a controlled demo is usually the easy part. The harder questions appear when the system has to handle real users, changing data, unreliable tools, production traffic, and decisions that may have real consequences.

This session looks at the issues engineers should think through before taking an agent into production. It will cover inference consistency, context limits, latency, cost, tool failures, security, human oversight, evaluation, observability, and recovery when something goes wrong.

It will also explore where MCP and A2A can be useful, what additional complexity they introduce, and why more autonomy or more agents does not always lead to a better system.

The goal is to share a practical way to evaluate these tradeoffs and build agent systems that remain reliable, secure, and maintainable after the demo ends.