Prime Agent is an open-source, self-improving coding agent harness from Prime Intellect built around two core abstractions: the Recursive Language Model (RLM), which treats context as a variable and sub-agent delegation as function calls inside a persistent IPython REPL, and the Continual Harness, which lets the agent CRUD its own prompts, skills, memory, and sub-agents at runtime. The harness supports persistent sub-agents, agent-to-agent messaging, autonomous eval mode with goal tracking and heartbeats, and a /refine pipeline that reads the agent's own trajectory to make incremental harness improvements. Using Opus 5, Prime Agent achieves 95.5% on ARC-AGI-3, surpassing the reported human expert baseline of 95.4%, while also outperforming native harnesses on long-context benchmarks including OOLONG, LongBenchPro, and EmulatorBench. Case studies include building emulators in Rust, writing GPU kernels, playing Factorio (including observed reward hacking), and navigating MazeBench. The project is fully open-source and installable via a single curl command.
