An agent that performs well in a demo still faces a harder test in production, where real users, changing prompts, and unstable tools expose hidden weaknesses. This session focuses on turning a working agent into a system you can trust. Using a single concrete agent as the running example, the session defines what reliability means for multi step, tool using behavior, including success, partial success, and failure modes. It then shows how to design evaluations that reflect real usage by building golden datasets grounded in actual user intent and scenario based tests that cover full action paths. You will also learn how to structure evaluation runs, score outcomes to surface brittleness and silent failures, and set up regression tests that detect breakage as prompts, tools, or APIs change over time.

What You will Learn

• How to define and classify reliability for multi step agents, including partial success and failure modes

• How to build production relevant evaluation suites using golden datasets and scenario based action path tests

• How to operationalize evaluation through scoring, regression testing, and monitoring for prompt, tool, and API drift

Who Should Attend

• Developers building or maintaining AI agents

• Engineers responsible for testing and reliability of AI systems

• AI and ML practitioners deploying agents into production environments

• Technical Leads overseeing quality and long term robustness of agent based systems