A new training method lets AI agents keep the skills of an elaborate support system even after that system is switched off.
Researchers describe Harness-Zero, a technique for what they call agent harness distillation. An agent harness is the external scaffolding (code, prompts, and tool wrappers) that mediates how a model reads its environment and takes actions, and specialized harnesses tend to boost performance only for the exact setup they were built for. Harness-Zero uses a highly tuned harness to guide a second "harnessing agent," which corrects a student model's responses so they work inside a simpler, fixed target harness. Those corrected trajectories become fine-tuning data, baking the specialized behavior into the model's weights so the elaborate harness can be removed entirely at deployment.
The gains hold up. Across knowledge-work, tool-use, and science tasks, the base model's average task success rose from 23.3% to 44.3% once the specialized harness was stripped away, actually beating the 41.7% it managed with that harness still attached. The model also recovered 82.3% of the specific behavioral patterns the harness had induced. That result echoes an old idea from a new angle: distillation has long been used to compress a big model's knowledge into a small one, but here the thing being distilled is not a model's outputs, it is the behavior of an entire tool-calling pipeline.
It is one paper's benchmark numbers, not a shipped product, and "internalizing behavior" is a tidy way of describing what is really a distillation shortcut around building better base models in the first place.