Object-Oriented Harness: Evolving Agent Harnesses through Object-Oriented Design
Abstract Overview
This paper presents OO-Harness, a framework for optimizing language-agent harnesses by evolving a hierarchy of executable classes using object-oriented design. Instead of relying on a single global harness, the method creates specialized subclasses through inheritance, method overrides, and added methods, while a router selects an appropriate harness for each task based on the task description, the base execution trajectory, and routing lessons collected during training. The framework evolves harnesses on discovery tasks with a fixed executor model and is evaluated on disjoint held-out tasks to assess generalization. The authors evaluate this approach across three long-horizon benchmarks covering business workflows, workspace tasks, and multi-tool task completion, accompanied by ablations on routing inputs and hierarchical inheritance.
Novelty
The core contribution is formulating harness optimization as object-oriented program evolution, where specialized harnesses are implemented as subclasses that reuse and extend parent harness code via inheritance. The framework pairs this hierarchical code evolution with a router that leverages accumulated routing lessons and observable base trajectories to determine whether to reuse the base harness or dispatch to a specialized class.
Results
Across AutomationBench, Workspace-Bench, and Toolathlon-GYM, OO-Harness achieves the highest primary scores among evaluated methods, including global harness optimization baselines, specialized harness families, and test-time scaling variants. It demonstrates up to a 15.79 percentage-point absolute improvement over the unevolved base harness, reaching 77.41% Mean on AutomationBench, 39.48% Mean on Workspace-Bench, and 37.45% PASS on Toolathlon-GYM. Ablation experiments confirm that both trajectory-informed routing and inheritance-based hierarchical structure contribute measurably to performance gains.
Key Points
- OO-Harness evolves a hierarchy of executable harness classes, allowing targeted execution strategies to be added through inheritance rather than maintaining a single global harness.
- At test time, a harness router uses task descriptions, observable base execution trajectories, class descriptions, and learned routing lessons to decide between reusing the base harness or selecting a specialized subclass.
- Empirical evaluations and ablations show that combining subclass specialization, code reuse, and trajectory-informed routing improves held-out task completion and preserves strong base performance on unfamiliar out-of-distribution tasks.