I'd let an AI agent search through training environments all day. I wouldn't let it spend my budget. That's the part of @datavendor_ai that makes sense to me. Agents can search private training data, compare listings, and check quality signals through MCP. If a taskset looks promising, you can even run a sample in your own HUD workspace before committing. But the actual purchase still happens in the web app. I think that's a better way to use agents. Let them handle the research and narrow down the options, while people stay in control of what gets bought.