Prime Intellect's Prime Agent Claims 95.5% on ARC-AGI-3 With Opus 5 — Just Past the 95.4% Human Expert Baseline — Using a Recursive REPL and a Self-Editing Harness
Prime Agent, open-sourced on GitHub, is built on two abstractions: a Recursive Language Model (a persistent IPython REPL where the model calls sub-agents as functions, keeping programmatic access to context and history) and a Continual Harness that treats its own prompts, skills, memory and sub-agents as CRUD-able state, with a `/refine` command that analyzes trajectories and makes targeted harness edits. Reported results: 95.5% RHAE Best@1 on ARC-AGI-3 with Claude Opus 5 against a 95.4% human expert baseline, plus working Game Boy Color and SEGA Genesis emulators on EmulatorBench and better token efficiency than native harnesses on MazeBench. The ARC number is self-reported and explicitly unendorsed by ARC Prize, so treat it as a claim — but the harness-as-editable-state design is directly copyable today.
↳ Follow the thread