Google researchers show how to drop the harness at runtime and boost performance

A new paper from Google and collaborators demonstrates that a method called Harness-Zero can distill the behavior of an optimized harness directly into the model itself, so that at deployment the model operates without any wrapper code and reaches 44.3 percent success on macro tasks, compared with 23.3 percent for a regular harness and 41.7 percent for the base model with a harness attached.

In this context a harness is the wrapper code that manages an agent’s action loop: it selects tools, parses outputs, handles errors and enforces a fixed action structure. The harness simplifies development and provides control, but it also constrains the action space the model can explore, adds overhead and makes it harder to move into environments where external code cannot run. The problem is well known: when the harness dictates the steps, the model learns to operate inside that frame rather than developing independent strategies.

Harness-Zero separates two action spaces. During training an enriched, optimized harness can correct the model’s responses before they execute. Those corrections become training demonstrations in the deployment action space, which is narrower. In other words, the complex harness serves only as a teacher during training; at inference the model runs alone, without wrapper code, but with the knowledge gained from those corrections.

Across 28 harness-induced behaviors spanning knowledge, tool use and science, the method achieves an average 82.3 percent recovery of the original behavior. For frontier models using the same evolved harness, the agent-as-harness variant outperforms the code-as-harness variant. The standout figure remains the 44.3 percent macro-task success without a harness at deployment versus 23.3 percent with a regular harness, a gap of more than 20 percentage points.

If the approach holds up, it opens a path to lighter-weight agents: no runtime dependency on a specific harness library, no need to maintain two separate codebases for training and deployment, and greater flexibility in constrained environments such as edge devices or secure settings that forbid running external code. The ability to transfer procedural knowledge from the harness into the model’s weights resembles classic knowledge distillation, only at the level of agentic behavior rather than logits.

Open questions remain. The paper (arXiv:2609.24974) is a pre-print that has not yet undergone peer review. The authors themselves note that robustness has not been deeply tested: it is unclear how the method behaves on out-of-distribution tasks, how many demonstrations each behavior requires, and whether the 82.3 percent average recovery masks wide variance across tasks. Until independent evaluations and open benchmarks appear, the results should be treated as an interesting proof of concept, not a production-ready solution.