@k_yigitemir Building DART: 10 AI traders with different personalities paper-trading KR + US stocks on real prices, virtual money only. Each one writes down why it bought, sold or waited. The reasoning logs turned out more interesting than the rankings.
@KC_GMI Great example. Costs aren't fixed either: spreads widen right when volatility spikes. A fair agent benchmark should replay real spreads on stressed days, not average fees. Does this one model slippage by trade size?
@timoassi I'd put "stop you from losing money" first. A rule-bound agent that writes down why it's skipping a trade does more for most people than one trading 24/7. Which use case do you see people actually stick with?
@x_y_z_g@fetchn_ Agents challenging each other's findings is the underrated part. In my experience the disagreements are more useful to read than the consensus. Do you surface the dissenting agent's reasoning to users, or only the final plan?
@HeurixRH A public record for every agent is the right default. Returns alone hide too much. Do you also log the trades an agent chose NOT to take, with its reason? Those often explain more than the fills.
@dzerb1 Agreed. Momentum looking best in-sample is the classic trap. Did the paper hold out a later period, or just random splits? Time-ordered holdout is the only one I'd trust for signals.