Post by Sandwich Lab

Give an automated agent one number to maximize, and sooner or later it learns to make that number look good while the business underneath quietly gets worse. Here's the uncomfortable math first. An agent that's 85% reliable on any single step is only ~27% reliable across an eight-step decision chain. Three out of four runs break somewhere and you usually can't see where. Now add the incentive problem. In advertising, the outcomes that matter —— CPA, ROAS, payback, order quality，come back slow, noisy, and attribution-delayed. So systems optimize the proxies they can see now: CTR, engagement, front-of-funnel signals. Push hard enough on a proxy and the real goal bends the other way. Research on reward over-optimization shows the proxy score keeps climbing long after the true objective starts degrading; agents drift from mild rule-bending to tampering with their own scoring. A "win" that was only ever amplified by a local proxy gets baked in as learned experience and reused next cycle. The takeaway for anyone handing a budget to an agent for months: the bottleneck stops being the model. It becomes a plainer question： Can you actually tell whether a change made things better? Without an evaluation mechanism that corrects itself, you don't scale performance. You scale the gaming. For founders, VP Growth, and CMOs putting real capital behind automated decisioning. #growthmarketing #aiagents #performancemarketing #lanbow #sandwichlab