CoworkOS:Scaling Agent Coworkers for Real-world Automation

CoworkOS is a cowork-agent harness that separates model-chosen task strategy from runtime execution management and supports long-horizon, human-in-the-loop work across benchmarks, real applications, and automated post-training.

Cowork agents must deliver high-quality artifacts on complex productivity tasks while also monitoring and making decisions over long horizons, a combination the paper calls an open challenge. The paper argues that today's cowork agents need both complex productivity task capability and continued monitoring and decision-making, with humans in the loop. It asks whether a single cowork agent can handle both while keeping humans in the loop.

CoworkOS divides responsibility: the model decides actions, delegation, and acceptance, while the runtime maintains state and feedback. It uses a three-layer runtime of Main Agent, Coordinator Runtime, and optional Subagents; Main plans and reviews, the Coordinator dispatches and monitors, and Subagents execute bounded assignments. A shared application layer lets humans and agents inspect and revise the same documents, presentations, CAD models, and Blender scenes, with a resource protocol that commits edits against a base revision and returns a conflict for stale proposals. Hierarchical workspaces and persistent semantic memory keep working files, decisions, and long-running Job logs shareable and resumable. In an RSI case study, CoworkOS uses GLM-5.3 as its backbone for automated post-training on Qwen3.6-35B-A3B.

• Across 12 benchmarks in four domains, CoworkOS ranks among the top two harnesses in every domain average, with Productivity 57.54, ML Coding 56.26, Tool Use 55.22, and Search 82.14. §4.1.2
• CoworkOS achieves the best results on SkillsBench, JobBench, WideSearch, MLE-Bench-Lite, and MLS-Bench. §4.1.2
• In the RSI post-training study, the average score over eight benchmarks rises from 33.19 to 42.95, a gain of 9.76. §4.3.3
• The final RSI model exceeds the initial baseline on seven of eight benchmarks and achieves the highest score across the six rounds on six, including ties. §4.3.3
• Task-level RSI gains include 33.15 points on Automation and 14.91 on SkillsBench over R0, while BFCL v4 gains 9.07 from R4 to R5. §4.3.3
• The qualitative comparison covers 24 cases and 72 artifacts across six categories. §A.2

• The paper states that RSI gains vary across tasks, with GDPVal and JobBench fluctuating overall and the best-performing round differing by task.
• The paper states that per-round RSI changes describe concurrent interventions rather than isolated causal effects, and that round index does not represent elapsed time.
• For MCP integrations that directly control external software, the paper states that the application retains its own document, saving, and concurrency semantics, and the platform does not claim guarantees the external application does not expose.
• The qualitative category analyses are stated to be grounded in representative observations from inspecting final deliverables, rather than reported as a quantitative artifact-quality score.

• CoworkOS separates task strategy from execution management: the model decides actions, delegation, and acceptance, while the runtime maintains state and feedback. §Abstract
• Main is the user-facing interface and overall planner, manager, and executor, and autonomously invokes Subagents for parallel processing or independent task handling. §3.1.1
• Communication flows through three directional paths: Main-to-Subagent, Subagent-to-Subagent, and Subagent-to-Main. §3.1.2
• The resource protocol treats identity and revision as first-class and commits an edit only when its base revision matches; otherwise it returns a conflict, and human and agent edits follow the same rule. §3.2.1
• In native applications, human edits enter the same coordination loop as agent actions, and the agent context includes the current resource and unacknowledged changes by other actors while excluding its own edits. §3.2.2
• MCP integration preserves an existing application's ownership and interface, while browser and desktop interaction is described as an orthogonal interaction mode rather than a third resource protocol. §3.2.3
• Agent working areas persist across Runs, each Run retains its own events, checkpoints, and results, and Subagents place reusable outputs in delivery areas for Main to review. §3.3.2
• Automated post-training combines complex delivery and sustained supervision: CoworkOS analyzes model weaknesses, optimizes the data pipeline, synthesizes and validates post-training data, feeds it into training, and iterates the closed loop. §4.3.1