Prime Intellect's Coding Harness Tops ARC-AGI-3 At 95.5%

Prime Intellect has launched Prime Agent, a general-purpose coding harness that has posted a score of 95.5 percent on ARC-AGI-3, a benchmark designed to test abstract reasoning and generalisation. This result puts the harness above the human-expert baseline on the test, a notable milestone given that ARC-AGI benchmarks are specifically constructed to resist brute-force pattern matching and reward genuine problem-solving ability. What makes this announcement significant is not just the raw score but the framing around it. Prime Intellect has stated that Prime Agent delivered major improvements across multiple underlying models when tested against the company's own previous proprietary harnesses. This suggests that the harness itself, the scaffolding and tooling that wraps around a model to help it reason, plan and execute tasks, is doing a substantial share of the work in driving performance gains. Rather than relying on a single flagship model to post record numbers, Prime Intellect's approach appears to boost results regardless of which model sits underneath it. This points to a broader trend in the AI coding and agent space: the harness, orchestration layer, or scaffolding is increasingly being treated as a first-class product rather than an afterthought. As foundation models converge in raw capability, the differentiation is shifting towards how well a system can structure tasks, manage context, retry failures, and coordinate multi-step reasoning. Prime Intellect's positioning of Prime Agent as a general-purpose tool, rather than one tuned narrowly for ARC-AGI-3, suggests the company is betting on this scaffolding layer as a durable source of competitive advantage. The announcement was made via Prime Intellect's official X account, with limited additional detail provided at this stage on which specific models were tested, the size of the improvement margins, or when Prime Agent might become more broadly available. As with many benchmark announcements in the current AI landscape, independent verification and real-world testing beyond the ARC-AGI-3 suite will be needed to confirm how well these gains translate to practical coding and agentic tasks outside the benchmark environment.