Agent Harnesses, Strategic Choice, and Environments Become Major Sources of Capability Gains
Prime Agent reportedly boosts ARC-AGI-3 from 30% to 95.5% Best@1, with harness choice producing 7.8x more score variance than model choice. Tasteful Agent, Skill2Env, RoboFoundry, EpiCon, AutoDataBench, Act First, Reason Later, and reproducibility studies show that strategic selection, reusable memory, environment design, data generation, and interaction scheduling are distinct bottlenecks. Weak transfer and evaluator overfitting remain major limits.
Sources (15)
Updated Sep 30, 2026