Prime Agent makes the harness persistent

Prime Intellect opened Prime Agent, a coding and research agent with a persistent IPython runtime. Its Recursive Language Model treats context as variables and subagents as programmatic calls; its Continual Harness stores supplemental prompts, memories, skills, and reusable subagent specifications as session-local state. Why it matters: Long-running agents accumulate more than chat history. Prime Agent makes the operating layer inspectable and refinable, while its documentation is clear about the tradeoff: model-generated Python runs with the worker's operating-system permissions.

Try this: Use a disposable clone for one bounded task. Save the diff, test output, and `prime-agent status` result, then inspect which harness state persisted and what generated code could access.