Continual Harness Released

Continual Harness is a reset-free adaptation loop in which, with only a minimal environment interface, the agent continuously updates its prompts, skills, memory, and sub-agents during execution.

In GPP (Gemini Plays Pokémon), which cleared Blue, Yellow Legacy Hard Mode, and Crystal without a loss, the harness editing that was previously done by humans increasingly shifted to being model-driven through meta tools such as define_agent, run_code, and notepad.

• In Pokemon Red/Emerald evaluations, it lowered button-input cost for frontier models and substantially narrowed the gap with a hand-crafted expert harness.
• With process-reward co-learning, a frontier teacher relabeled rollouts from an open-source agent to update the model, and it showed continued progress on Pokemon Red with no restarts.

The paper's message is that the performance of long-horizon agents depends not on the model alone, but on the self-improvement of the harness itself.