From general intelligence to production-grade intelligence
General models learn what is common. Production ships what is rare.
Four parts work continuously in parallel. Agent harness runs are recorded as trajectory data and scored by domain-specific benchmarks. Trajectories and benchmark judgments feed training. Training updates the model weights; the improved model returns to the agent harness.
Together they form a system that learns continually from production: first in the harness, then in the model weights.