openJiuwen: Beyond Static Harnesses for Long-Horizon Coding Agents
The openJiuwen Team (with collaborators across Singapore and China) release an open-source coding-agent harness built around two named design goals, Structural Composability and Runtime Adaptivity, and report SWE-bench Verified and Terminal-Bench 2.1 numbers above the best official leaderboard points.
Ask this paper
Harness, not model: The model policy is held fixed and every gain comes from framework-controlled runtime decisions, which makes this a clean argument that the harness is the remaining lever on long-horizon coding.
Rail composition: Capabilities are composed as Rails over a shared execution substrate that spans single agents, delegated sub-agents, and a Swarm Flow mode, so developers reconfigure orchestration without rebuilding it per task.
Evidence-driven control: Semantic diagnostics, execution outcomes, task progress, and context relevance are fed back as runtime evidence that reshapes context, feedback, and task control mid-run rather than at design time.
Numbers: 82.6% on SWE-bench Verified and 87.19% on Terminal-Bench 2.1, ahead of the strongest selected official-leaderboard point estimates by 3.4 and 3.39 points.
Why it matters: Two harness properties, composability for the developer and adaptivity for the run, are stated as first-class requirements rather than incidental engineering, which is a useful frame for anyone building agent infrastructure.
Abstract
Long-horizon coding agents operate over evolving repository states while increasingly relying on heterogeneous capabilities, delegated agents, and multi-agent coordination. These trends pose two complementary challenges for the agent harness. First, developers need to compose capabilities, reconfigure execution logic, and scale increasingly complex agent systems without repeatedly rebuilding orchestration. Second, complex coding tasks continuously produce new evidence