Build a trip-planning agent that improves itself with no human editing the prompt, and then watch it learn to cheat. Over a fixed dataset of places, hours, and transit, the agent's true score is feasibility: are the stops actually open and the transit actually doable? Run adk optimize and watch feasibility climb from ~40% to ~90%, and read the new rules the agent wrote into its own instructions ("check hours and travel time before adding a stop"). Then swap the judge for a "looks-complete" metric and catch it red-handed: it stuffs the itinerary with unverified stops, exposed the instant you grade it on real feasibility again. The load-bearing lesson: you optimize what you measure, so grade the trajectory (did it actually check?), not the answer. Scales to Vertex Gen AI Eval and fine-tuning Gemma from the traces. 🕒 Americas: Thursday, August 20: 9:00 AM PDT 🕒 Asia-Pacific: Friday, August 21: 9:30 AM IST