Prime Agent: open-source self-improving AI coding system hits 95.5% ARC-AGI-3

Prime Intellect, a distributed AI infrastructure startup valued at $1B after its Series A, released Prime Agent—an open-source, self-improving coding harness built on a Recursive Language Model (RLM) framework. Prime Agent is designed to let AI agents write code, run tests, learn from results, and adjust behavior without a human in the loop. Technically, Prime Agent uses a Continual Harness running inside a persistent Python REPL so coding context carries over step to step. In that environment, context is treated as a living variable. The model can call tools, delegate tasks to sub-agents, and update its approach based on outcomes, reducing reliance on static prompts. On performance, Prime Agent uses Opus 5 as its underlying model and scored 95.5% on the ARC-AGI-3 benchmark, surpassing the human-expert baselines the test is calibrated against. The launch follows Prime Intellect’s deployment of its verifiers stack (version one) and the creation of hundreds of thousands of sandboxed environments via its Environment Hub. From a market perspective, Prime Intellect is positioning Prime Agent as an infrastructure layer rather than a locked-down model product. It claims startups and enterprises can build on it without licensing fees or vendor lock-in, aligning with venture capital’s shift from foundation-model winners toward tooling and production infrastructure. For traders, the headline is AI infrastructure momentum rather than direct token catalysts—watch for sentiment spillover into AI-themed crypto sectors.