Harnesses for Inference-Time Alignment over Execution Trajectories

In three linesStudy of harness engineering for inference-time alignment of LLM agents. Authors decompose harnesses into task decomposition and guided execution mechanisms. They identify failure modes (over-decomposition, over-pruning) and show partial harnesses specifying only initial steps can outperform fully structured workflows.

Summary generated by Claude — human-verified