SimGym¶
Shopify's customer simulation environment. Browser-based agents play buyer and merchant roles, driven by years of real Shopify session data (not just prompts). Used to test theme changes, UI experiments, notification flows, and buyer-side interventions before they go to real merchants.
The insight Parakhin keeps returning to: if you prompt an agent in a vacuum to "be a customer," it will do exactly what you prompted — giving you a simulation that tautologically confirms your hypothesis. With historical behavior data as a calibration signal, agents act more like the population they're supposed to represent. This is the basis of simulation-data-moat.
Underneath: SimGym uses hstu (Hierarchical Sequential Transduction Units) to model both buyer trajectories and merchant-as-entity trajectories, enabling counterfactual rollouts from intervention nodes ("what if we sent this buyer a 10% coupon at minute 3 of the session?").