Action Scaling at the Harness Boundary Beats Trajectory Re-Runs

A new research paper from NVIDIA and KAIST, 'Mid-Harness: Scaling Actions Between Model and Harness for Terminal Agents,' proposes a more efficient way to improve AI agent performance. Instead of relying on expensive 'trajectory scaling'—where entire multi-turn sessions are re-run to fix errors—the authors suggest 'action scaling' at the model-harness boundary. By generating multiple candidate bash commands and using a verifier to select the best one before execution, agents can avoid compounding errors that contaminate the environment. The study found that a distilled 9B parameter model using pairwise verification achieved higher success rates than brute-force trajectory re-runs at a fraction of the compute cost. Furthermore, the researchers discovered that single-token logit-based verification outperforms chain-of-thought reasoning, as it avoids narrative drift and hallucinations. This approach offers a significant shift in how developers can optimize agent reliability while reducing inference expenses.