Prime Intellect Open-Sources Prime Agent, Hitting 95.5% on ARC-AGI 3

Prime Intellect open-sourced Prime Agent on August 25, followed by a technical report the next day. It is a self-improving RLM (reasoning language model) framework for coding and long-horizon autonomous tasks, best known for scoring 95.5% on ARC-AGI 3—surpassing the human expert baseline of 95.4%—a leap from the previous 30% level. According to @aigclink, the framework runs stably and saves compute.

Confirmed

Prime Agent is open source (confirmed by Prime Intellect's official account, m5), using recursive sub-agents, persistent compute, and inter-agent coordination to extend long-horizon capabilities
A core mechanism is a persistent REPL: a persistent IPython kernel replaces conversation turns, with variables surviving across turns so the model can programmatically manage its own context
The technical report (relayed by @rohanpaulai) introduces a memory hierarchy spanning IPython sessions and disk storage, letting the model move state across storage tiers via code rather than relying solely on context compression—turning long-context work into an information-management problem
It scored 95.5% on ARC-AGI 3, above the human expert score of 95.4%

Why it matters

95.5% vs 95.4% marks a crossover above the human expert baseline, a jump from 30%
Methodologically, it shifts long-horizon agent work from "bigger context windows" to "managing state with code," offering the community a reusable open-source paradigm

2026-08-25 ~ 2026-08-26 · 6 related posts

Primary sources

Prime Agent: An Open-Source Harness for Self-Improving RLM — PrimeIntellect · 2026-08-25 11:17 (UTC+08:00)
Prime Agent technical report: memory hierarchy enables state management via code — rohanpaul_ai · 2026-08-26 06:39 (UTC+08:00)
Prime Agent: Open-Source Framework Boosts ARC-AGI Score from 30% to 95.5% — dair_ai · 2026-08-25 22:39 (UTC+08:00)

[source] Prime Agent: An Open-Source Harness for Self-Improving RLM — PrimeIntellect · 2026-08-25
Prime Agent report: Self-improving RLM harness for long-horizon tasks — amaarora · 2026-08-25
[source] Prime Agent: Open-Source Framework Boosts ARC-AGI Score from 30% to 95.5% — dair_ai · 2026-08-25
[source] Prime Agent technical report: memory hierarchy enables state management via code — rohanpaul_ai · 2026-08-26
Prime Agent framework tops ARC-AGI 3 with REPL-based programming flow — aigclink · 2026-08-26

1 near-duplicate retellings: aigclink