Prime Agent

Rather than bolting an agent loop onto a chat API, Prime Agent implements the Recursive Language Model pattern: the model works inside a persistent IPython kernel where context lives in variables, tools are ordinary function calls, and an rlm() call spawns real child agents for parallel or background work. Prime Intellect released the MIT-licensed harness in August 2026, and it collected 5,200 GitHub stars within two days on its way past 18,000. The continual-learning angle is the distinctive part: supplemental prompts, memories, and skill descriptions are durable state the agent itself can edit, and a /refine command reviews the current trajectory and applies small, evidence-backed updates, so the harness sharpens with use instead of resetting every session. Daemon-backed sessions keep long-running jobs alive after the terminal disconnects. Installation is a one-line shell script on macOS or Linux, and authentication goes through model provider API keys or subscription logins for Claude, ChatGPT, or GitHub Copilot, so a connection to an inference endpoint is assumed throughout.