HomeNews

CoworkOS Beats Codex Across 12 Agent Benchmarks

Saurabh Anand
Saurabh Anand
•
Nov 3, 2026 3:45 PM
•
0
 min read
Select Emergent as your Preferred news source
CoworkOS Beats Codex Across 12 Agent Benchmarks

💡 TL;DR

  • CoworkOS, a three-layer agent harness from Renmin University and five partners, claims top scores on four of 12 benchmarks against Codex and DeepSeek.
  • RSI post-training lifted Qwen3.6-35B-A3B's average across eight benchmarks by 29.4 percent, from 33.19 to 42.95 over five rounds.
  • The architecture separates strategy from execution, enabling human-agent collaboration on evolving documents, presentations, CAD models and Blender scenes.

A new agent harness from Renmin University of China and five academic partners claims performance competitive with or ahead of Codex, OpenHands and DeepSeek Harness across 12 benchmarks spanning tool use, productivity, search and machine-learning coding. Officially released on November 3, 2026, CoworkOS places first on SkillsBench (64.34), JobBench (59.88), WideSearch (82.47), MLE-Bench-Lite (76.90) and MLS-Bench (35.62), according to a preprint published alongside the system.

Three-Layer Architecture Separates Strategy and Execution

CoworkOS introduces what its authors describe as a three-layer design: the model layer decides actions, delegation and acceptance; the runtime layer maintains state and feedback; and humans and agents collaborate on the same evolving artifacts. The paper emphasizes "long-horizon work" scenarios in which both human operators and automated agents inspect and revise documents, presentations, CAD models and Blender scenes. This separation of concerns is intended to let strategy and execution scale independently, a pattern that recent agent research has explored to reduce context limits and improve reliability.

RSI Post-Training Lifts Qwen3.6-35B-A3B by 29.4 Percent

The headline experiment uses GLM-5.3 as the backbone to conduct five rounds of RSI-driven automated post-training on Qwen3.6-35B-A3B. Over those five rounds, the model's average score across eight benchmarks rose from 33.19 to 42.95, a gain of 9.76 points or 29.4 percent. The preprint does not break down which individual benchmarks contributed to the aggregate lift, but the result suggests that iterative self-improvement loops can deliver measurable gains when anchored by a capable teacher model.

Benchmark Scores Against Codex and DeepSeek Harness

CoworkOS was evaluated on 12 benchmarks across six application categories: tool use, productivity, search, machine-learning coding, general coding and reasoning. The system reported top scores on five benchmarks:

  • SkillsBench: 64.34
  • JobBench: 59.88
  • WideSearch: 82.47
  • MLE-Bench-Lite: 76.90
  • MLS-Bench: 35.62

The authors describe these results as "highly competitive" with Codex and DeepSeek Harness, though head-to-head comparisons on all 12 benchmarks were not published in the abstract. The breadth of categories tested suggests the harness is designed for general-purpose agent work rather than a single vertical.

Collaborators Span China and Singapore

The preprint lists co-authors from Fudan University, Peking University, Shanghai Jiao Tong University, the National University of Singapore and East China Normal University alongside Renmin University of China. The collaboration arrives amid a surge of agent-harness research: tracking data shows 526 agent-related stories published in the last 90 days, indicating sustained academic and industry interest in frameworks that orchestrate model calls, tool use and human feedback at scale.

What This Means

CoworkOS adds another data point to the evolving landscape of agent orchestration, particularly for tasks that span multiple tools and require iterative refinement. The RSI training loop demonstrates that post-training on agent-specific tasks can yield double-digit percentage gains, a pattern likely to inform future harness designs. Whether the system's claimed edge over Codex and DeepSeek Harness holds across all 12 benchmarks will depend on full evaluation details, but the multi-institution collaboration and breadth of tested scenarios signal confidence in the architecture's generality.

About the writer

Saurabh Anand Rai is the Head of Product at Emergent, where he leads product strategy and innovation for AI-powered tools that help people build software faster.

Start Building
on Emergent today
Try Emergent