Prime Agent lifts ARC-AGI-3 to 95.5%

Prime Intellect’s open-source Prime Agent harness raises ARC-AGI-3 RHAE Best@1 from 30% to 95.5% by adding a persistent Python environment, durable context, recursive subagents, and explicit recovery and verification, according to the paper and code release. The result is a harness comparison, not a claim that the underlying model suddenly became 3.2 times smarter; teams evaluating coding agents should version orchestration and context management beside model weights, as the archive’s same-model harness cost analysis argues.