Prime Intellect open-sourced Prime Agent, a coding harness built around an architectural idea that's been circulating in agent research but rarely shipped in a usable tool: instead of giving a model a fixed menu of tool-calling functions and a static system prompt, give it one tool — a persistent Python REPL — and let everything else, including spawning other agents, happen as code running inside that REPL. Under this Recursive Language Model design, calling a sub-agent looks like a normal async function call rather than a special-cased tool-calling schema, which means the model can compose sub-agent calls with ordinary control flow: loops, conditionals, parallel fan-out, and results stored in variables it can inspect later. Because the REPL persists for the whole session, the agent doesn't lose access to earlier results the way context-window compaction normally forces it to — old data lives in a variable, not in a chat transcript that eventually gets summarized away. The second piece, which Prime Intellect calls the Continual Harness, lets the agent edit its own scaffolding while it works: it can create, update or delete its own memories, skills, sub-agent definitions and prompt notes through the same interface it uses for any other data structure, and a refine step periodically looks at what just happened in a task and applies the smallest useful edit to that scaffolding — for example, turning a repeated failure into a saved note, or promoting a workaround into a reusable skill. The results are notable less for beating other harnesses on generic benchmarks (though Prime Agent using Claude Opus 5 scored 95.5% on ARC-AGI-3, edging past the cited human-expert baseline) and more as a demonstration that harness design, not just model scale, is still a live lever for agent capability. It's MIT-licensed and works with subscription logins or API keys across most major model providers, so teams building their own agent harnesses have a concrete, runnable reference implementation of the pattern rather than just a paper describing it.