AI-agent failures may show when to switch from harness tweaks to weight training

Researchers tested a self-evolving harness—the tools and runtime around a frozen model—on DeepPlanning’s multi-day travel tasks. They report that it raised Qwen3.5-4B’s held-out score from 0.16 to 0.30. Training adapters on evolved-harness trajectories added 0.13 to held-out scores under the original harness for both model sizes. The researchers propose diagnosing failures first: use harness changes for process problems and weight training when completed plans are poor.