An open-source agent harness just surpassed the human expert ceiling on the hardest AI reasoning benchmark. It did it by giving the model a programming interface to its own context.

The Summary

  • Prime Agent scored 95.5% on ARC-AGI-3 with Claude Opus 5 — above the human expert baseline — and outperforms proprietary harnesses across multiple frontier models
  • The key innovation: a persistent IPython kernel as the model's only tool, letting it program over its own history, launch subagents, and manage state outside the active context window
  • Self-improvement happens mid-task via /refine — not just after completion like Hermes — rewriting prompts, memory, and skills in real time
  • Subagents persist after finishing and can message each other directly, enabling genuine multi-agent coordination without a central orchestrator
  • Fully open-source, open license, model-agnostic — tested with Opus 5, GPT-5.6 Sol, GLM, and Kimi

The Signal

On August 5, Prime Intellect dropped Prime Agent — a self-improving RLM (Recursive Language Model) harness for coding and long-running autonomous tasks. The architecture is built on three ideas stacked together: programmatic tool calling through a persistent IPython kernel, persistent multi-agent orchestration where subagents survive task completion and communicate peer-to-peer, and a Continual Harness that lets the model rewrite its own scaffolding while work is still in progress.

That last part is the piece that matters. Every existing harness — Hermes, Claude Code, the proprietary wrappers — improves the agent between tasks. Prime Agent does it during. The /refine call lets the model examine its own prompt, memory, and tool definitions mid-execution and overwrite them. The context becomes a variable the model can program over, not a fixed container it fills up. Long sessions become a programming problem rather than a context-management problem.

The benchmark results are not toy numbers. On ARC-AGI-3, widely regarded as the most rigorous test of genuine reasoning rather than pattern completion, Prime Agent with Opus 5 scored 95.5% — above the human expert baseline. On EmulatorBench it built SEGA Genesis and Game Boy Color emulators from scratch in Rust, passing diagnostic tests for hardware accuracy. On Factorio and MazeBench — long-horizon planning tasks with no ceiling — it outperformed existing harnesses by wide margins. The gains held across open-weight models too: GLM and Kimi showed the same pattern, meaning the architecture is doing real work, not just amplifying an already-strong closed model.

Token efficiency is the counterintuitive part. By using a single IPython kernel as the interface rather than a sprawling tool registry, the model makes fewer, denser calls. It manages its own context rather than having the harness manage it. The result is fewer tokens spent for more capable behavior — the opposite of the usual tradeoff when you add scaffolding.

The Implication

The distinction between "harness" and "agent" has been blurring for two years. Prime Agent collapses it. When a model can rewrite the instructions it runs on, manage its own memory architecture, and spawn persistent subagents that collaborate without a coordinator — you are no longer describing a tool with a wrapper. You are describing a system that modifies itself toward a goal. That is a qualitatively different thing to deploy, and a qualitatively different thing to align.

The open-source release is significant in both directions. On one hand, it means any team can now run agents that match or exceed what proprietary labs are shipping in their native harnesses. On the other, the safety surface of self-modifying agents running on open-weight models with no usage policy is genuinely new territory. The benchmark win on ARC-AGI-3 is impressive. The more important test is what these systems do in production when the task is ambiguous, the context is adversarial, and the model decides mid-run to rewrite what it was told to do. We do not have good benchmarks for that yet. We are about to need them.

Sources

Prime Intellect on X — Prime Agent launch thread | Prime Agent GitHub (pi) | Travis Wright on X