RUC & Fudan release CoworkOS: agent runtime for long-horizon work with human edits, boosting benchmarks from 33.19 to 42.95

jiqizhixin · x · 2026-11-20

Renmin University of China and Fudan University released the official technical report for CoworkOS, targeting the bottleneck where complex artifact tasks outlast a single conversation and humans make edits along the way.

Mechanism: the model decides actions and acceptance while a runtime maintains execution state; a coordinator manages delegated work, shared application objects let humans and agents revise the same artifacts, and persistent memory plus job logs enable resumption.

Results: in a sustained-work experiment, a GLM-5.3 agent ran five rounds of automated post-training on Qwen3.6-35B-A3B, raising its 8-benchmark average from 33.19 to 42.95 (AutomationBench +33.15pp, SkillsBench +14.91pp, etc.). CoworkOS also leads OpenHands, Codex, and DSH across SkillsBench, Tool Use, JobBench and other benchmarks.

Original post →

More from coding & agent

coding & agent channel →