An Open Problem in RL: Beyond Verifiable Rewards

One dilemma in RL today is that models can be trained successfully mainly on tasks such as math and code, where there is a clear logical notion of right and wrong, an exact reward signal, and a formal answer. But how should we apply RL to tasks such as research, agent planning, and writing?

For research-like tasks, part of the current solution is implemented through an agent harness. Can the research-related part of that harness be integrated into the model's own intelligence? Or should the rest of the harness, including the capabilities needed to complete long-horizon tasks, be written as a complete system, analogous to Internet networking protocols, so that it can be applied across different models?