Harness-Zero: PKU, Google and HKUST Distill Agent Harnesses into Model Weights
AxSaucedo · x · 2026-09-28
Researchers from Peking University, Google, and HKUST introduce Harness-Zero, a paper on agent harness distillation.
Problem: External harnesses boost agent performance, but gains stay tied to the harness at deployment. The best harness varies by domain, instance, and model, so a general-purpose agent must settle for a suboptimal shared harness or route among many specialized ones.
Method: Use a domain- or instance-optimized harness as training-time guidance. A harnessing agent corrects student responses in the target harness's action space before execution, converting harness guidance into training demonstrations. Fine-tuning on these trajectories internalizes harness-induced behavior into the weights, so the specialized harness can be removed at deployment.
Findings (across knowledge work, tool use, and science domains):
- For frontier LLMs with the same evolved harness, agent-as-harness outperforms code-as-harness;
- After distillation, models retain the specialized harness's gains even with it removed.
More from coding & agent
- Claude Code creator Boris Cherny: bet on general models, skip fine-tuning — rohanpaul_ai · 2026-09-28
- Personal agents need fixed chores, not more tokens: define boundaries before letting them run — sujingshen · 2026-09-28
- App devs face two paths in the personal-agent era: integrate MCP or become the agent — sujingshen · 2026-09-28
- 30 verified Claude Opus 5.5 browser animation cases, ranked by views, with prompts — dotey · 2026-09-28
- One agent per household: whose veto wins when family calendars conflict? — sujingshen · 2026-09-28
- Rolldown to ship experimental inlineCommonChunks to cut small shared chunks — cnakazawa · 2026-09-28