Before expanding from a low-volume pilot to the entire catalog, Wayfair scientists needed to significantly reduce the API costs of the new LLM-based product information validation system. Using AI-agents, the science team supercharged the pace of experimentation, running hundreds of experiments over two days, exercising 40+ unique design levers, most of which were newly thought of and implemented. The winning configuration reduced cost by 95%, exceeding expectations. The other, equally important, big win was that an effort that would have traditionally required 2-months of planning and execution, with many ideas left untried, was compressed into 2 days with no ideas left unexplored. Three months later, with another AI-accelerated research push, the team produced a further 95% cost reduction on the same product information validation system. This framework is gaining traction at Wayfair and has been applied to more research problems. Science teams are spending their time doing the fun stuff – whiteboarding, coming up with interesting ideas – and having AI agents do the grunt work. After each application of our AI-accelerated research framework, we have gathered our learnings and improved our practices. For example, for traditional ML research, you establish separate development and evaluation sets. In the era of agentic coding, you’ll want to keep the evaluation set well hidden behind a metric reporting API, to avoid reward-hacking-induced overfitting.