Deploying agentic AI in production introduces a unique engineering challenge: debugging non-deterministic execution paths. Unlike traditional software, an agent's "code" is a dynamic interplay of prompt context, model weights, and external tool outputs. This talk presents a rigorous engineering methodology for agents, focusing on quantifying agent performance and analysing trajectories. Key takeways: →Trajectory evaluations: Analysing the "Reasoning Trace" using a secondary judge model to detect hallucinated logic steps or tool misuse. → Cost-latency trade-offs: Optimizing token usage via dynamic context compression and speculative execution of tool calls. → Sandboxing & side-effects: Technical implementation of ephemeral execution environments to safely contain agentic code execution.