I think most AI trading agents need a skeptic agent, not another strategy agent

It’s getting ridiculously easy to have an LLM write a trading strategy, turn it into code and backtest it. The output often looks legit too. Clean logic, decent equity curve, a reasonable explanation for why it should work. But it can still be nonsense because of one tiny assumption: using info that wasn’t available yet, entering at an impossible price or trying enough variations until one happens to look good. The agent doesn’t need to be “wrong” for this to happen. It just needs to be very good at producing plausible stuff fast. I’m more interested in agents that try to break strategies than agents that generate them. Something that checks timing, data leakage, execution assumptions and whether the result survives out-of-sample data. Has anyone here built something like that?