CoworkOS claims parity with Codex, DeepSeek Harness

CoworkOS arrives against a backdrop of intense activity in agent-harness research; the write-up notes 526 agent stories in the last 90 days on the tracking site's side. The paper's framing centers on "long-horizon work" — tasks that stretch beyond a single step — and its three-layer architecture is meant to keep strategy and execution apart, with humans and agents inspecting and revising the same files, including evolving documents, presentations, CAD models and Blender scenes.

The headline experiment is a self-improvement run: GLM-5.3 served as backbone for "RSI-driven automated post-training" on Qwen3.6-35B-A3B, and over five rounds the model's average score across eight benchmarks rose from 33.19 to 42.95 (+9.76). The abstract itself provides no per-benchmark breakdown. The co-author list spans Fudan, Peking, Shanghai Jiao Tong, NUS and East China Normal alongside Renmin, suggesting a multi-institution effort rather than a single-lab release.