PrimeIntellect open-sources Prime Agent, scores 95.5% on ARC-AGI-3, beating human experts

PrimeIntellect announced the fully open-source release of Prime Agent, a self-improving agent framework designed for coding and long-horizon tasks. The framework achieved a score of 95.5% on the ARC-AGI-3 benchmark, surpassing the human expert baseline, and the developers stated it was not specifically optimized for this benchmark. Its core design converts long-horizon AI sessions into programming problems, leveraging a persistent IPython kernel as context, combined with recursive language models (RLM), persistent multi-agent orchestration, and a continual harness for self-improvement, achieving high token efficiency and self-evolution. The project quickly gained attention, with GitHub stars rising from 1.1k to over 5,000.

Confirmed

• Prime Agent scored 95.5% on ARC-AGI-3, surpassing the human expert baseline. According to @virusxp, the framework improved Opus 5's ARC-AGI-3 score from 30.2% to 95.5%.
• The developers stated the framework was not specifically optimized for any benchmark and showed significant improvements over native tools across models.
• The core design philosophy is to convert long-horizon AI sessions into programming problems, primarily relying on a persistent IPython kernel as context. It combines three core technologies: recursive language models (RLM) with native programmatic tool calling, persistent multi-agent orchestration, and a self-improving continual harness.
• The framework is highly token-efficient, supporting context as a variable, allowing the model to programmatically search history, call tools, spawn sub-agents, and self-evolve based on its own context and test environment.
• Practical tests show that although initially designed for coding tasks, it also excels at general agent tasks in any domain, performing well on challenging long-horizon tasks like Factorio and MazeBench.
• The project is open-sourced on GitHub; according to @willccbb, it gained 1.1k stars shortly after release, and later @PrimeIntellect-ai showed over 5,000 stars.

Why it matters

• Prime Agent demonstrates the feasibility of model self-evolution through recursive models and a continual harness, breaking the context rot limitation of traditional agent architectures, and providing a new architectural approach for solving complex long-horizon tasks.

2026-08-06 ~ 2026-08-07 · 40 related posts

• Prime Agent Launches: Token-Efficient Self-Improving Harness for Coding Agents — latkins · 2026-08-06
• Prime Agent Coding Harness Tops ARC-AGI-3 with 95.5% Beating Human Experts — xeophon · 2026-08-06
• Introducing Prime Agent: A Self-Improving Framework for Coding and Autonomous Tasks — a1zhang · 2026-08-06
• Prime Agent: Fusing Recursive LLMs with Self-Improving Orchestration — dejavucoder · 2026-08-06
• Prime Agent Achieves SOTA on Complex Long-Horizon Tasks Like Factorio — xeophon · 2026-08-06
• PrimeIntellect open-sources Prime Agent: a self-improving coding agent — TheZachMueller · 2026-08-06
• AI Agent 'Prime Agent' Scores 95.5% on ARC-AGI-3 Benchmark — ___Patrice___ · 2026-08-06
• Prime Intellect Releases Self-Improving Agent, Beating Human Baseline on ARC-AGI-3 — mattbeane · 2026-08-06
• Prime Agent Coding Harness Surpasses Human Experts on ARC-AGI-3 — lifebypixels · 2026-08-06
• PrimeIntellect Launches Prime Agent: An Open-Source Self-Improving Coding Agent — andersonbcdefg · 2026-08-06
• Prime Agent Framework Massively Boosts ARC-AGI Performance — daniel_mac8 · 2026-08-06
• [source] Prime Agent Scores 95% on ARC-AGI-3 Benchmark Using Opus 5 — virusxp · 2026-08-06
• PrimeAgent faces backlash: ARC-AGI-3 high score accused of overfitting — 机器之心 · 2026-08-07
• Prime Intellect Ships Open-Source Prime Agent to Solve LLM Context Rot — blaizedsouza · 2026-08-07
• [source] Prime Agent: A Self-Improving RLM Agent for Coding Workflows — PrimeIntellect-ai · 2026-08-07
• PrimeIntellect Open-Sources Prime Agent: Self-Improving Agent That Evolves Without Changing Weights — aigclink · 2026-08-07
• Why AI Harnesses Matter More Than Models: Deep Dive into Prime Agent — Prompt Engineering · 2026-08-07

23 near-duplicate retellings: CShorten30 · CShorten30 · ricklamers · lateinteraction · willccbb · ResearchCrafty1804 · omarabudayyeh · iamrobotbear · generativist · willccbb · generativist · willccbb · omarabudayyeh · lateinteraction · burny_tech · willccbb · khademinori · ChrisGPT · thomasahle · afurgs · TheZachMueller · xiaohu · omarsar0