Prime Intellect Releases Prime Agent, a Self-Improving Harness That Hits 95.5% on ARC-AGI-3

Prime Intellect on August 5, 2026 unveiled Prime Agent, an open-source, self-improving agent harness for coding and long-running autonomous work that scored 95.5% Best@1 on the interactive ARC-AGI-3 benchmark—edging past a human-expert baseline of 95.4%. The result is notable because frontier models running on their own have historically scored under 1% on ARC-AGI-3, underscoring how much of the gain comes from the scaffolding layer rather than the underlying model.