The research indicates that by using a custom harness with memory handling and a "supervisor" component, AI agents can perform significantly better, even if the underlying model is not optimized for the task. For example, Claude Opus 5 scored 100% on the interactive reasoning benchmark ARC-AGI-3 with the harness, compared to 30% without it. This highlights the importance of the agent's scaffolding, which manages memory, context, and feedback, in improving AI performance and cost efficiency for complex, multi-step operations.