Harness RL is Meta-Learning: Training to Self-Improve at Test Time

Self-improving agents must learn to update themselves at test time. Meta-learning trains exactly such update rules by optimizing post-adaptation performance, but because its inner loop runs in model weights, it is prohibitively costly at LLM scale. A modern policy, however, is more than weights: it is a frozen model plus a harness of instructions, memory, and tool-use rules. We observe that relocating the inner loop to the harness removes the computational obstacle: harness revision needs no gradients, so the meta-objective becomes a regular RL objective on post-adaptation return. We introduce harness RL: a proposer is trained to revise the harness from observed rollouts while the executor stays frozen, so every gain must flow through the written harness. We test harness RL on reasoning and coding tasks and observe both generalization and test-time scaling. What harness RL improves is thus not performance on a fixed task but the model's ability to improve itself, a capability we call meta-self-improvement.