November 1 agent guide: Mage-VL-Cowork hybrid computer use

November 1, 2026 is a computer-use day for Bacancy Technology readers, not a frontier chat-model day. Microsoft Mage-VL-Cowork is a 4B hybrid agent that decides when to click and when to code. The Microsoft Mage Team reports 66.0% on OSWorld-Verified under a 50-step budget, with Apache 2.0 weights on Hugging Face. DeepSeek Super-agent remained the October 31 research prototype, and the November 1 design essay about that prototype is not a second product launch.

What shipped on November 1

Mage-VL-Cowork: GUI and code in one policy

Mage-VL-Cowork is a multimodal hybrid computer-use agent built on Mage-ViT plus Qwen3-4B. The model card lists 11 GUI primitives in normalized coordinates, a persistent bash and Python channel, a 32,768-token active window, and a working memory that keeps the three most recent screenshots with folded action history. That hybrid design is the point: many computer-use agents stay stuck in mouse and keyboard loops when a short shell command would finish the job.

Reported results and training shape

• 66.0% on OSWorld-Verified (two-run average) under a 50-step budget at 4B parameters.
• Online hybrid RL adds 10.16 percentage points over Hybrid SFT while cutting mean model calls by 13.0% and output tokens by 25.4%, per Microsoft.
• Mage-CUA-Forge synthesizes 293,806 desktop and web tasks across 6,915 skills for the training curriculum.
• Category scores reported by Microsoft include 87.5% on OS tasks and 62.7% on multi-application workflow tasks.

Treat those scores as vendor-reported until an independent OSWorld run lands. The technical report on arXiv (2610.40747) and the Hugging Face card are the primary references. The release matters for teams that want an inspectable, self-hostable computer-use policy instead of only a closed hosted harness.

Why it matters for developers and businesses

Developers get a concrete 4B baseline for hybrid desktop agents: when to ground on a screenshot, when to open a shell, and how to keep memory bounded. Businesses get a smaller on-device or private-cloud option for repetitive desktop work where sending every frame to a frontier API is not acceptable. Bacancy Technology recommends a narrow pilot: pick ten internal desktop tasks, run Mage-VL-Cowork and the current default agent side by side, and compare completion rate, step count, and human takeover rate.

What this day is not

• Not a new Claude, GPT, or Gemini frontier chat model.
• Not a second DeepSeek Super-agent product. The November 1 essay points at the October 31 research prototype: /blog/october-31-deepseek-superagent-claude-code-2026.
• Not a replacement for managed coding agents such as Claude Code or GitHub Copilot.

Pilot steps

1. 01
   Pull the Apache 2.0 checkpoint

   Load microsoft/Mage-VL-Cowork from Hugging Face and follow the Mage project harness notes before writing a custom loop.
2. 02
   Define a 50-step budget and success criteria

   Match the published budget so the pilot is comparable. Record completion, human takeover, and unsafe actions separately.
3. 03
   Keep the agent off production desktops until the eval passes

   Hybrid bash access can change files and call network tools. Sandbox first.

Conclusion

November 1 is a hybrid computer-use release day. Mage-VL-Cowork gives Bacancy Technology clients a small, open policy to bench against larger closed agents. Keep Super-agent and Mage in different buckets: one is a multi-agent isolation pattern, the other is a desktop CUA checkpoint.

Related reading

• October 31 Super-agent and Claude policy: /blog/october-31-deepseek-superagent-claude-code-2026
• October 30 Claude Code write blocks: /blog/october-30-claude-code-write-blocks-2026
• Agentic AI topic map: /topics/agentic-ai
• Multi-agent systems: /topics/multi-agent-systems

Need a computer-use pilot designed against your desktop workflows?

Bacancy Technology designs sandboxed computer-use evaluations for IT and product teams. Consulting: /consulting/ . Contact: /contact/ .

Sources: Mage-VL-Cowork model card, Hugging Face . Project page, microsoft.github.io — cowork . Technical report, arXiv paper .