Prime Agent is an open-source agent framework built around a persistent IPython REPL that implements the “Recursive Language Model” abstraction, letting recursive subagents share history, memory, and skills across long-horizon trajectories. By standardizing execution, verification, and resource management inside the harness rather than the underlying model, the framework isolates harness limitations from model capability. In testing, Prime Agent raised ARC-AGI-3 scores from 30% to 95.5% and matched or exceeded existing solutions on coding, GPU-kernel generation, and complex simulation tasks such as Factorio.
