According to Dynalysis Beating monitoring, the AI training platform Prime Intellect has open-sourced Prime Agent. It is a runtime framework deployed outside the model, responsible for maintaining history, invoking tools, scheduling sub-agents, and able to adjust its prompts, memories, and skills based on failure records. The model parameters remain unchanged. The project is released under the MIT license.
Prime Intellect ran the ARC-AGI-3 public tasks three times using Opus 5, achieving scores of 95.0%, 95.2%, and 95.5% respectively. The highest score slightly surpasses its referenced 95.4% human expert baseline.
However, it is not a community first. The ARC Prize community leaderboard already includes Tycho with 100%, Retrodict with 99.9%, and baseline1 with 99.0%. These solutions also fall under the category of "model plus custom runtime framework," leveraging memory, code execution, or a world model to assist AI in problem-solving.
