A preprint presents an open-source coding-agent harness that scored above selected systems on two coding benchmarks. With GPT-5.6 Sol, openJiuwen achieved 87.19% accuracy on Terminal-Bench 2.1, 3.39 percentage points above the selected strongest official-leaderboard result at 83.8%. On SWE-bench Verified, it resolved 82.6% of 500 instances, 3.4 percentage points above the selected strongest leaderboard result at 79.2%. The paper describes these as system-level comparisons, so the margins do not by themselves show how much came from the harness.
A framework built to adapt
The paper presents openJiuwen as an open-source harness designed to make capabilities easier for developers to combine and to adapt execution across long, multi-step coding tasks. It uses a shared execution substrate and Rail-based capability composition across a single agent, delegated sub-agents and Swarm Flow.
The runtime design adapts the framework's choices about context, feedback and task control around a fixed model policy as the evidence from a task changes. In practical terms, the surrounding workflow can respond to new task information while the underlying model policy stays fixed.
Two tests, different yardsticks
OpenJiuwen was evaluated on two collections. SWE-bench Verified supplied 500 human-validated software-engineering tasks derived from real-world GitHub issues. Terminal-Bench 2.1 supplied 89 diverse tasks run in containerized terminal environments. The yardsticks were different: SWE-bench used Pass@1, the percentage of instances successfully resolved, while Terminal-Bench used accuracy from task-specific verifiers.
For SWE-bench Verified, the authors used Claude Opus 4.5 with high reasoning effort. For the primary Terminal-Bench 2.1 result, they used GPT-5.6 Sol and also evaluated Fable 5 for model-matched comparisons. Within each openJiuwen configuration, the prompt, available tools and model configuration remained fixed, and outcomes were graded by official test suites or task-specific verifiers.
Where the numbers landed
On Terminal-Bench 2.1, openJiuwen achieved 87.19% accuracy with GPT-5.6 Sol, compared with 83.8% for the selected strongest official-leaderboard result. The difference was 3.39 percentage points. The reported uncertainty values were plus or minus 1.20% for openJiuwen and plus or minus 1.2% for the selected result.
In the Fable 5 model-matched comparison, openJiuwen achieved 84.04% accuracy, versus 83.8% for Claude Code and 80.4% for Terminus 2. The reported point estimates put openJiuwen 0.24 percentage points ahead of Claude Code.
That comparison also broke results out by task category. On file operations, the scores were 0.76 for openJiuwen, 0.56 for Claude Code and 0.52 for Terminus 2. On system administration, they were 0.889, 0.778 and 0.844, respectively. Those category gaps are not a controlled test of the harness because tool availability was not isolated.
On SWE-bench Verified, success varied by estimated fix duration. Using Claude Opus 4.5 with high reasoning effort, openJiuwen reported success rates of 91.75% for tasks estimated at under 15 minutes, 81.23% for tasks estimated at 15 minutes to 1 hour, and 52.38% for tasks estimated at 1 to 4 hours. The authors did not draw a strong conclusion from the more-than-four-hour bucket because it contained only three tasks.
What the comparisons can show
The main qualification is that the benchmark margins do not isolate openJiuwen's contribution. The authors characterize the cross-system comparisons as system-level results because leaderboard systems may use different models, prompts, tools and agent implementations. That means the study does not establish that the harness alone caused the differences.
The evidence is also narrow in coverage. The current evaluation focuses on SWE-bench Verified and Terminal-Bench 2.1, and the authors call for broader controlled studies across additional benchmarks, models and configurations, with more detailed ablations. For now, the reported advantage belongs to the tested configurations, not to a standalone measurement of the harness.
Paper data and sources
Original title: openJiuwen: Beyond Static Harnesses for Long-Horizon Coding Agents
Authors: openJiuwen Team, Tao Yu, Xinyu Zhang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text