Prime Agent Boosts ARC-AGI-3 Benchmark from 30% to 95.5%

burkov · x · 2026-08-27

A paper from MIT, PrimeIntellect, and Princeton introduces Prime Agent, an open-source multi-agent execution framework. It combines persistent programmatic environments, recursive sub-agent communication, and cross-trajectory memory to scale long-horizon problem-solving. The framework boosts performance on the ARC-AGI-3 benchmark from 30% to 95.5%.

Related event: Prime Intellect Releases Prime Agent, Boosting ARC-AGI-3 to 95.5%(3 posts)→

Original post →

More from coding & agent

coding & agent channel →