Prime Agent, the harness that put Opus 5 at 95.5% on ARC-AGI-3, now has a technical report.

The mechanism is a memory hierarchy. Model weights and active context sit underneath a persistent IPython session and a disk-backed store of histories, skills and prompts, and the model moves state between those levels with code instead of having it compacted away.

Long inputs stay in the REPL as variables the agent can search and transform, so long-context work becomes an information-management problem rather than a reading problem.

– arxiv. org/abs/2608.23552

Title: "Prime Agent: A Self-Improving RLM Harness"