openJiuwen: static harnesses lose when long-horizon evidence keeps arriving
Aug 28 openJiuwen team: Rail-composed Inner/Outer loops + runtime adaptivity — SWE-bench Verified 82.6% (+3.4 pp) and Terminal-Bench 2.1 87.19% (+3.39 pp) over strongest selected leaderboard points.
Source: openJiuwen: Beyond Static Harnesses for Long-Horizon Coding Agents
Huawei’s openJiuwen team names two harness failures continuous factories hit as soon as agents run past a single session: Structural Composability (capabilities and topologies accumulate as bespoke orchestration) and Runtime Adaptivity (pre-run decisions go stale once diagnostics, tests, progress, and context pressure change mid-trajectory). Their open harness answers both with a shared Inner Loop/Outer Loop substrate, Rail-based lifecycle hooks that compose the same way for single agents, delegated sub-agents, and Swarm Flow, and framework-controlled adaptation around a fixed model policy — Context Management, Goal Mode (acceptance vs resource stop), LSP-driven passive feedback, and cross-task Self-Reflection.
On SWE-bench Verified with Claude Opus 4.5 (high), openJiuwen resolves 82.6%, +3.4 pp over the strongest selected non-bash-only leaderboard point (79.2%, also Opus 4.5). On Terminal-Bench 2.1 it posts 87.19% with GPT-5.6 Sol (+3.39 pp over Claude Code + Fable 5 at 83.8%); model-matched Fable 5 reaches 84.04%. Duration buckets on SWE stay competitive into the 1–4 hour band (52.38% vs 35.71% for mini-SWE-agent at the same Opus high setting), consistent with context management under long trajectories rather than just a hotter model.
Code: openJiuwen-ai/jiuwenswarm. Caveats they own: system-level leaderboard comparisons, not fully controlled harness-only ablations; Goal Mode still task-level; Self-Reflection and offline evolution are complementary, not yet one loop.
Why it matters for code review
Always-on factories that ship a static harness and hope the night’s evidence will be ignored are optimizing for the first hour. Review overnight control the way openJiuwen splits the problem: compose capabilities on one execution substrate, and let newly arrived tests, LSP diagnostics, and goal assessments change context and stop — without swapping the model mid-run. Continuous verified momentum compounds when morning merge gates see a harness that adapted under volume; FOMO for idle agents is wasted if the fleet kept replaying a plan written before the first failing check.