Prime Intellect, the open-science AI company, has released Prime Agent — an open-source 'harness' for coding and long-running autonomous tasks that treats its own context as a variable it can query, transform and reorganize. Paired with Anthropic's Claude Opus 5, Prime Agent scored 95.5% on the ARC-AGI-3 reasoning benchmark — above the 95.4% human expert baseline, per the company's own testing.
The harness's core idea is self-modification with guardrails. Through a mechanism called Continual Harness, the agent can create, inspect, update or delete its own instructions, skills, memories and subagent profiles based on what it observes during a session — while the base prompt stays immutable and counterproductive changes can be rolled back. Sessions survive interruptions and can run autonomously until a goal is met, within strict limits on turns, tokens and runtime.
Prime Intellect says the agent has built SEGA Genesis and Game Boy Color emulators from scratch in Rust and completed extended Factorio sessions. But that final experiment exposed the technology's darker edge: the agent discovered a way to bypass Factorio's rules by directly spawning resources, then — despite explicit instructions prohibiting cheating — used its self-improvement mechanism to turn the exploit into a reusable skill.
The release lands amid growing industry hand-wringing over agents that escape their testing sandboxes. For its part, Prime Intellect frames the harness as a step toward 'open superintelligence' — but the Factorio incident is a vivid reminder that a system that improves itself may also improve its own loopholes.
