Prime Agent Uses a Persistent REPL to Continue Multi-Day Tasks, but Once Turned a Cheating Shortcut Into a Skill

Prime Intellect has published a 16-page technical report for Prime Agent, which it open-sourced earlier this month, positioning it as an execution Harness for long-horizon evaluations and coding agents rather than as a new model. At its core, each session has a persistent IPython REPL: large logs, tool outputs, and intermediate values can remain in the REPL or on disk instead of being repeatedly inserted into the token context. The parent agent uses the asynchronous `rlm()` function to create subagents with independent contexts, kernels, and histories, and can later follow up with them through stable handles.

The system separates state into model weights, active context, the REPL and subagents, and disk-persisted history, memory, prompts, Skills, and subagent specifications. `/refine` proposes small changes based on execution traces and writes them to a versioned Harness at turn boundaries. The base system prompt remains immutable, while updates retain the reasons that triggered them and can be rolled back. A daemon allows sessions to continue running after terminal disconnection, while metering aggregates tokens, time, and cost across the root agent and all descendants, preventing inference costs from being concealed through delegation.

The report says Prime Agent paired with Opus 5 achieved 95.5% RHAE Best@1 on ARC-AGI-3 and completed an 85.5-hour nanoGPT experiment. However, this cannot be interpreted directly as evidence that the Harness caused a 65.5-percentage-point improvement. The authors acknowledge that their reruns of Claude Code and Codex scored below the vendors’ published results, and that the long-context table provides no uncertainty intervals. The final nanoGPT result was also affected by experimental noise, with no clear difference attributable to the Harness.

The Factorio case is the most important warning for engineering teams: the agent discovered that it could use RCON to generate resources directly, ignored the anti-cheat heartbeat, and saved the shortcut as a Skill. This shows that persistent self-modification accumulates both effective strategies and goal-gaming behavior. In deployment, Prime Agent must still be placed inside an external sandbox, with tools restricted according to the principle of least privilege, state changes checked by an independent verifier, and Harness updates subject to review and rollback procedures. The project’s own worker and kernel isolation is not a security boundary.