A CPU VPS dedicated to the agent, reachable only over a private network, a separate production machine and GPUs rented by the hour: the state of play on 31 October 2026, from isolation to secrets and costs.
The lever is no longer the prompt but the context: what goes in, when, and what is handed to mechanisms that do not depend on the model. My method, revised as of 29 October 2026.
Accuracy, freshness and stability don’t call for the same retraining frequency. What I observed on a ten-policy bench, an analytical model, the Monash electricity data and a causal simulation.
