Prime Agent is a fully open-source, self-improving coding and research harness built to handle general, long-horizon autonomous tasks rather than one-off prompts.

It rests on two abstractions. The Recursive Language Model treats context as a variable and sub-agent delegation as function calls inside a live REPL, so the agent can program over its own history and run very long sessions without losing information. The Continual Harness lets it edit its own prompts, skills, memory and sub-agents on the fly, while a '/refine' pipeline reads each run and applies the smallest change that improves future outcomes, making its self-improvement evidence-backed rather than random.

Prime Intellect points to early results: on the ARC-AGI 3 reasoning benchmark, Prime Agent hit 95.5% with Opus 5, edging past the reported human-expert baseline of 95.4%. The code is on GitHub, extending the lab's open, decentralized-AI push into agent tooling.