Open-source harness hits 82.6% on SWE-bench Verified

An open-source agent harness called openJiuwen reports 82.6% on SWE-bench Verified and 87.19% on Terminal-Bench 2.1, beating the strongest leaderboard entries it compared against by 3.4 and 3.39 points.

The numbers matter less than how the harness gets there. Most agent frameworks fix a plan up front and execute it. openJiuwen splits the problem in two. One layer, which the team calls Structural Composability, lets developers compose single agents, delegated sub-agents, and a "Swarm Flow" of multiple agents over one shared execution layer. The other, Runtime Adaptivity, feeds semantic diagnostics, execution outcomes, and task progress back into the framework so it can adjust context, feedback, and task control while the task runs, instead of sticking to a script written before the agent saw any real output.

Agent benchmarks have climbed for a year on bigger models and better prompts. This result comes from restructuring how the harness reacts to its own execution, with the model choice held fixed.

What sets this apart from a lot of benchmark papers is that the code is running, not shelved. The org's jiuwenswarm repo has 6,309 stars and its agent-core repo has 420, both pushed the same day as the paper. That's a real difference from an arxiv result with no implementation to check. The paper doesn't say how the harness handles a task where execution feedback is noisy or contradictory, which is the exact case Runtime Adaptivity is supposed to solve.

For anyone building multi-agent systems, read the composability piece first. Single agents and sub-agents sharing one execution layer is the structural change the runtime-adaptivity claims depend on.