Prime Agent is a new self-improving coding harness built on two core abstractions: the Recursive Language Model (RLM) and the Continual Harness. Unlike older designs that force models to work around static scaffolding, Prime Agent leverages frontier model capabilities by treating context as a variable and subagent delegation as function calls within a persistent REPL.

This design grants the model programmatic access to its history, subagents, and tools, enabling it to write language model programs that operate on its own context. The Continual Harness allows the agent to dynamically manage its own state, including prompts, skills, and memory, enabling sophisticated orchestration and agent-to-agent communication. This allows for the creation of persistent subagents that can be messaged later in a trajectory or even across different Prime Agent sessions.

Prime Agent is designed for use with both open and closed frontier models and is fully open-source. It functions as a general coding assistant, a runtime for autonomous evaluation, and a collaborator for research. The system utilizes an IPython kernel as its primary tool, with subagents implemented as separate Prime Agent instances. A background daemon manages all live sessions, ensuring recoverability even if a worker process crashes.