PrimeIntellect has released Prime Agent, an open-source coding and research agent designed for general and long-running autonomous tasks. The system is built on pi and features a self-improving reinforcement learning mechanism with programmatic tool calling and multi-agent messaging.

  • It achieves a 95.5% score on ARC-AGI-3, surpassing the human-expert baseline.
  • The harness demonstrates major improvements across models compared to their proprietary counterparts.
  • Prime Agent is fully open-source under an open license.

The release provides developers with a token-efficient and expressive framework for complex coding workflows.