Prime Intellect Open-Sources Prime Agent: Self-Improving Coding Harness With Recursive Language Models and 95.5% ARC-AGI-3 Score

Prime Intellect has open-sourced Prime Agent, a self-improving coding harness built on the Recursive Language Model (RLM) abstraction where sub-agents run as function calls inside a persistent IPython kernel. The framework allows agents to create, read, update, and delete their own state—prompts, sub-agents, skills, and memory—during execution. With Anthropic's Opus 5, Prime Agent achieves 95.5% on ARC-AGI-3, exceeding reported human expert baselines. The harness includes background daemon session management, agent-to-agent messaging scoped to prevent cross-session chatter, and automatic recovery from crashes using session logs and kernel snapshots. Prime Intellect reports lower token usage than native harnesses and demonstrates practical applications including building Rust emulators from specifications and optimizing GPU kernels over 1,000+ tool calls in 24-hour runs.