Prime Intellect's Prime Agent hits 95.5% on ARC-AGI-3, beating humans

Prime Intellect has released Prime Agent, a new open-source harness designed for coding and long-running autonomous research tasks. It's a self-improving system built on their pi model, and on the ARC-AGI-3 reasoning benchmark, it scores 95.5%—surpassing the human-expert baseline and outclassing proprietary harnesses like OpenAI's Codex and Claude Code. The team emphasizes that the gains aren't benchmark-specific; they see consistent improvements across multiple models when paired with Prime Agent compared to their default setups.

The secret lies in its architecture: programmatic tool calling lets the agent execute code directly, while treating context as a variable enables flexible memory management across long tasks. Multi-agent messaging allows multiple agent instances to collaborate, and the harness state is self-modifiable, meaning it can adjust its own workflow mid-task. All of this makes it both token-efficient and expressive. Fully open-source with an open license, Prime Agent is available on GitHub, giving any developer a production-ready alternative to expensive, closed coding agents.