When developing Agent applications, choosing a Harness is often treated as choosing a universal optimum. The report "Finding the Right Fit: Model-Harness Interactions across Agent Tasks," released by Professor An Bo's team at Nanyang Technological University in Singapore, gives the opposite answer: the quality of a Harness is not a fixed property but depends on its specific combination with the model and the task.
The experimental design itself is the premise of this conclusion. The study selects four Harnesses that can freely configure models—OpenHands, DSH (DeepSeek Harness), PI, and openJiuwen—and connects them respectively to five models: Claude Opus 5, GPT-6 Astra, GLM-5.3, Kimi K3, and DeepSeek V4 Pro, then places them on three task sets with different orientations: TUA-Bench focuses on general terminal operations, ALE-CLI on professional workflows, and Terminal-Bench 4 contains more difficult command-line tasks.
The 4x5x3 design yields 60 results, and adding the 6 results from two native pairings—Codex-GPT and Claude Code-Claude—across the three task sets, a comparison matrix of 66 records is ultimately formed. Because the three task sets differ in task composition and scoring criteria, the study does not merge the results into a single composite score but compares them separately.
Scoring uses fixed task-set denominators: TUA-Bench has 120 tasks, ALE-CLI has 99 tasks, and Terminal-Bench 4 has 63 tasks in total; the first two allow partial credit, while Terminal-Bench 4 is scored as pass or fail. Tasks interrupted by infrastructure failures were rerun, each task was scored only once based on its last run, and tasks without valid results were recorded as 0 points.
The run configuration was unified as much as possible: all Harnesses call models through OpenRouter and are fixed to route to each model's official provider, reasoning effort is uniformly set to high, and context management, retries, and turn limits remain at each Harness's default settings. TUA-Bench and Terminal-Bench 4 were run using Harbor v0.22.0, ALE-CLI used the task set's own runner, and all Harnesses were connected to the 14 computer operation tools provided by the task sets.
It is worth noting the comparison boundary: Codex and Claude Code only connect to their own models and did not participate in the full crossed experiment, so the comparison between native pairings and configurable Harnesses is not entirely equivalent. The scope of the conclusions is also limited to these four Harnesses, five models, and three task sets.