Harness-Zero Moves Agent Harness Behavior Into a Distilled Model
O’Reilly Radar reports that Haoran Ye and Peking University colleagues distilled an evolved agent harness into Qwen3.5-9B, then removed the harness and reviewer during testing. The distilled model’s average task success rose 21 points over its untrained predecessor, exceeding the improvement obtained by adding the full evolved harness, with evaluations covering spreadsheet editing, multi-app tool use and chemistry. The report says only generalizable harness behavior can move into the model; application-specific tools, data, users and environments still require a harness.
The result suggests that some hand-built agent scaffolding can become redundant as models absorb learned behavior, while shifting ongoing engineering work toward testing, maintaining and rebuilding application-specific harness components.
This source provides an excerpt. Continue reading on the original site.
#Harness-Zero#Peking University#Qwen3.5-9B#Kimi K3#GPT-5.6 Sol