Alibaba's DreamX lifts Qwen 3.7-Plus to 80.7% on WeaveBench

A paper out of Alibaba this week is worth pausing on, because it points at where quick agent gains might still hide. In an arXiv preprint indexed on Hugging Face, the DreamX Team describes a scaffolding layer called LongHorizon-Harness that, without any retraining, lifts Qwen 3.7-Plus from 51.8% to 80.7% on WeaveBench, from 69.7% to 77.2% on Terminal-Bench 2.1, and from 2.8% to 8.3% on OSWorld 2.0. Applied to Claude Opus 4.7 on a 34-task OSWorld 2.0 subset, the same wrapper reportedly moves the model from 20.0% to 34.3%.

The design idea is simple. Instead of one long agent session that plans, acts, and grades itself inside the same growing context, the harness splits the work across three roles in a Manage-Execute-Audit loop. A manager holds the task state outside the execution loop and picks the next subtask. A fresh-context executor performs only that subtask. A read-only auditor then independently inspects the environment before anything is written back to state. Only audit-verified facts survive across rounds; the executor's raw trajectory is discarded each round. The paper reframes long-horizon execution as a task-state management problem rather than a context-length problem.

Why this matters if you are not building agents from scratch: it suggests that a lot of the gap between today's models and reliable long-horizon work sits in orchestration, not raw model capability. If you are choosing between waiting for the next frontier model and investing in the wrapper around your current one, this result is an argument for the wrapper. The paper describes the AgentAdapter as backend-interchangeable across Claude Code, Codex CLI, OpenClaw, and Hermes Agent, so the pattern is not tied to a single vendor.

The honest caveats are worth flagging. These numbers come from a single team reporting on its own system, and the Qwen 3.7-Plus WeaveBench runs used root privileges inside the task virtual machine while the official baselines used a regular user account, which the authors themselves treat as reference points rather than matched comparisons. The paper also does not price the compute cost of running three roles across up to 25 rounds per task, with the executor budgeted at 1800 seconds and the manager and auditor at 300 seconds each. What the reporting does not give you is an independent audit-failure rate or performance outside these three benchmarks. Still, the direction is the part worth watching: harness design is starting to matter as much as model size for long-horizon work.