Prime Agent Released

Prime Intellect has released Prime Agent, an open-source agent harness for long-running autonomous tasks and coding, under the MIT license.

Prime Agent uses Recursive Language Models (RLMs) and a Continual Harness as its core abstractions. The agent is designed to treat conversation history and task results as variables, allowing it to directly modify prompts, skills, memory, and sub-agent specifications during execution.

The tools exposed to the model are unified into a single persistent IPython kernel, covering file manipulation, shell execution, search, and sub-agent invocation. The workflow proceeds by having the model write code, the kernel execute it, and the model read the results.

However, the entire system is not composed solely of Python. A TypeScript host handles provider calls, session persistence, scheduling, child agent lifecycle, and safety policies, while IPython serves as the programming interface interacting with the model.

It can be installed with a single-line script on macOS and Linux, and can be combined with Prime Intellect's prime-rl and verifiers to serve as an agent training environment and evaluation runtime.