Prime Intellect has released Prime Agent, an open-source self-improving agent framework under the MIT license. In a post on its official blog, the company said the framework, running on Opus 5, scored 95.5% on the ARC-AGI-3 benchmark, slightly above the 95.4% human expert baseline. The company framed the result as a gain delivered by redesigning the external harness around existing models rather than training a new model from scratch.
Prime Intellect said ARC-AGI-3 initially saw every frontier model score below 1%, making the jump to above-human performance in a matter of months notable in its view. It also said the framework improved results across multiple models relative to their model-specific harnesses, not only on this single test. Prime Agent is built around two ideas: a Recursive Language Model, which treats context as a programmable variable and turns sub-agent delegation into function calls in a REPL environment, and a Continual Harness, which allows the agent to add, read, modify, and delete its own prompts, skills, memory, and sub-agents. The company added that the framework can also be used as an outer-layer harness for Claude Code and Codex.
Prime Intellect has open-sourced Prime Agent, a self-improving agent framework released under the MIT license. In its official blog, the company said that with Opus 5 as the underlying driver, Prime Agent scored 95.5% on ARC-AGI-3, edging past the 95.4% human expert baseline.
A framework change, not a new model
Prime Intellect said the significance of the result lies in how ARC-AGI-3 started out: when the test was first introduced, every frontier model scored below 1%. The move from under 1% to above the human expert line took only a few months, according to the company, and came from redesigning the harness wrapped around existing models rather than training a new model.
The company also said the framework produced clear gains across multiple models compared with each model's own dedicated harness, suggesting the effect was not limited to a single benchmark run.
Two core components
Prime Agent is built on two concepts. The first is a Recursive Language Model, or RLM, which treats context as a programmable variable and turns delegated sub-agents into function calls inside a REPL environment.
The second is a Continual Harness. Under that design, the framework's own state — including prompts, skills, memory, and sub-agents — becomes something the agent can add, read, modify, and delete on its own.
Prime Intellect said the framework can also serve directly as an outer-layer harness for Claude Code and Codex.
This article was originally published by Bit.Fan. For more cryptocurrency news and market insights, visit www.bit.fan. Disclaimer:
The market information, project data, and third-party content displayed on this platform are for industry information sharing only and do not constitute any form of investment advice or return commitment.
Cryptocurrency trading carries high risks. Users should fully assess their risk tolerance and make independent decisions. All profits, losses, and legal responsibilities are borne by the users themselves.
