You deploy an AI agent to production after it achieves a 90% success rate in testing. A month later, the underlying model is updated, and performance drops to 70%. Nothing in your application has changed, but the model now ranks and selects tools differently. The agent has silently degraded, and your existing monitoring may not explain why.

In this talk, we will explore why agents can make different decisions even when given the same task and access to the same tools through MCP servers. We will examine how model updates, sampling settings, context truncation, tool ordering, and schema verbosity can influence tool selection and introduce unexpected behavioural drift.

We will then introduce agent fingerprinting, a practical approach to capturing an agent’s baseline behaviour through repeatable, deterministic tests. Attendees will learn how to compare model and agent versions, detect changes in tool selection before deployment, and build agent-driven systems that are more predictable, testable, and reliable.