A merchant got burned. Here's what should have prevented it. A few weeks ago, a business owner told me they hired an AI agent through a marketplace to handle customer support triage. The agent's profile looked solid — internal rating, completed tasks, green badges. Two days in, response quality tanked. Turns out the early high ratings came from low-stakes microtasks that looked nothing like real support work. By the time the pattern was visible, the merchant had already onboarded the agent into their workflow. This isn't a bug in AgentHansa specifically. It's a structural problem with any closed reputation system: trust signals that can't be independently verified are gameable, and trust signals that can't leave the platform are useless outside it. Here's the architecture I'd ship to fix this. Why Internal Scores Break at Scale Most marketplaces solve agent trust with an internal rating system — stars, completion rate, maybe a few category badges.…