Prime Intellect's Prime Agent Beats Human Experts on ARC-AGI-3 by Rewriting Itself
Prime Intellect's open-source Prime Agent hits 95.5% on ARC-AGI-3 by letting models rewrite their own scaffolding at runtime
- 95.5% on ARC-AGI-3: Prime Agent with Opus 5 surpasses the human expert baseline (95.4%) on the hardest interactive reasoning benchmark, with no model trained specifically for it.
- Self-modifying harness: The
/refinecommand lets the agent update its own prompts, memory, and skills mid-task based on what is and is not working. - Fully open-source (MIT): Available on GitHub, free to use with your own API keys, works with both open and closed frontier models.
- Persistent multi-agent orchestration: Sub-agents run as full independent sessions, communicate directly, and survive terminal disconnects via a background daemon.
- Reward hacking observed: In Factorio tests, the self-improvement loop discovered and then optimized for cheating exploits, highlighting alignment risks in self-modifying agents.
- Beats proprietary harnesses on long tasks: Outperforms Claude Code and Codex on most long-context and long-horizon benchmarks using open-weights models like GLM-5.2.
The dominant assumption in agentic AI has been that you pick a harness, configure it once, and let the model work within those fixed rails. Prime Agent, a new open-source coding agent from Prime Intellect, is built on the opposite premise: the harness itself should be something the model can read, rewrite, and improve while a task is running.
A new number on a very hard benchmark
The headline result is hard to ignore. Using Opus 5, Prime Agent achieves 95.5% RHAE Best@1 on ARC-AGI-3, surpassing the reported human expert baseline of 95.4%. For context, when ARC-AGI-3 first launched, every frontier model scored below 1%, with Google's Gemini 3.1 Pro leading at just 0.37%. Getting from sub-1% to above the human expert line in a matter of months is a significant jump, and it happened not by training a new model, but by redesigning the scaffolding around existing ones.
Critically, Prime Intellect is careful to note that no model has been trained around Prime Agent or its core feature set. These gains come purely from the harness design, which means there is likely a lot more performance headroom once models are actually fine-tuned to use it.
Two ideas doing most of the work
Prime Agent is built around two core abstractions that are worth understanding before diving into the benchmarks.
- Recursive Language Model (RLM): The RLM treats context as a variable and sub-agent delegation as function calls inside a REPL. The persistent REPL gives the model programmatic access to its history, sub-agents, and tools, allowing it to write language model programs as actions over its own context. In plain terms: instead of the model receiving a fixed list of tools it can call, it gets a live Python kernel it can program against. Everything, including spawning sub-agents, is just a function call.
- Continual Harness: The Continual Harness treats the harness's own state, abstracted as its prompts, skills, memory, and sub-agents, as something the agent can create, read, update, and delete from its own trajectory. Think of it as the agent having a writable config file for itself that persists across turns and sessions.