Continual Harness: An Efficient Self-Improving Agent for ARC-AGI-3

ARC-AGI-3 is an IQ test that measures an agent's intelligence, requiring the agent to form an Internal World Model that understands rules and mechanisms through new evidence.

Continual Harness is a reset-free self-improving agent harness that supports memory storage, writing reusable skills, deploying sub-agents, and self-modifying prompts. This approach is designed to work even in extreme environments with no prior information about the game's rules or tools.

The key results are as follows:

• High Efficiency: On the ARC-AGI-3 public set, it scored 20.54% at a cost of just $774, overwhelming the existing Hermes (8.25%, $5,674) and OpenClaw (5.20%, $2,912).
• Skill Reuse: 62% of all executed actions came from Saved Skills rather than new reasoning. This exceeds 80% in top-performing games.
• Action Optimization: While Hermes used 15.30x the human baseline in actions per completed level, Continual Harness achieved its goals with only 1.48x the actions.

In conclusion, because useful computational results are integrated as part of the harness state rather than ending as one-off tasks, Continual Harness dramatically boosts performance by building on knowledge gained during the initial exploration phase.