Peking University team trains Qwen3.5-9B to absorb agent harness behavior
Overview
A Peking University research team trained Qwen3.5-9B on corrected runs produced by an agent harness, then ran the model without the harness, according to an O'Reilly Radar article.
The distilled model's average task success rose 21 points over its untrained predecessor and beat the model-plus-harness combination. The article's author argues that harnesses are still needed for tools, data, users and the environment, and that teams should tag traces with harness and model versions and disable harness components one at a time when a new model ships.
Written by AI from the articles below · updated Oct 27, 7:30 AM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- O'Reilly RadarBlogAgent harnesses are changing, not dying, as Harness-Zero trains harness behavior into a model
AIA Peking University research team trained a small model, Qwen3.5-9B, on corrected runs produced by a harness, then removed the harness. The distilled model's average task success rose 21 points over its untrained predecessor, beating the model-plus-harness combination. The article argues that harnesses remain necessary for tools, data, users, and environment, and that teams should tag traces with harness and model versions, then disable components one at a time when a new model ships.
Heat Trend
- Current comparable heat
- 5
- Peak comparable heat
- 10Oct 27, 8:00 AM ET
- Change in past 24 hours
- –
The trend compares only continuously observed participants, so its scope may be smaller than the current heat total. Hover or tap for hourly heat; use arrow keys with a keyboard.