Evals are the new QA — knowing your agent is still good after the model changes

AI agents can impress in a demo, but the real challenge is ensuring they continue to deliver reliable, measurable outcomes as models, prompts and tools evolve. As organisations move AI agents into production, traditional software testing is no longer enough. Non-deterministic outputs make it harder to define success, identify regressions and confidently introduce changes without compromising performance.

This roundtable will explore how teams can build robust evaluation frameworks that turn agent quality into something measurable, repeatable and actionable. Discussion will examine why defining “good” is a product decision requiring clear ownership, agreed success criteria and alignment between product, engineering and domain experts. Attendees will also consider how production traces, real-world failures and human interventions can inform representative evaluation datasets that reflect genuine user needs.

We will explore the opportunities and limitations of LLM-as-judge approaches, including the importance of calibrating automated assessments against human judgement. Participants will discuss how model upgrades can be managed through structured regression testing, controlled rollouts and rollback strategies, enabling teams to adopt new models with greater confidence.

Finally, the conversation will examine how reliable evaluations can connect technical performance to business outcomes, including outcome-based pricing. With AWS services such as Amazon Bedrock model evaluation and Amazon Bedrock AgentCore Evaluations and Observability, organisations can explore how to operationalise evaluation, improve visibility and scale AI agents with confidence.

Amazon Web Services (AWS) is the world's most comprehensive and broadly adopted cloud platform, offering over 200 fully featured services from global data centers. Millions of customers, including the fastest-growing startups, largest enterprises, and leading government agencies, leverage AWS to reduce costs, enhance agility, and accelerate innovation. Since 2006, we've transformed industries and lives by working backwards from our customers' needs to provide tailored cloud infrastructure solutions.