Copilot computer use and Qwen3.8-Omni-Flash: desktop agents and omnimodal workers in early October 2026

Introduction

Coding agents spent 2026 gaining MCP, cloud sandboxes, and coordinator projects. Two surfaces still leaked work into humans: legacy desktop apps and long audio-video jobs. Early October closed the first gap in Copilot. Late September closed more of the second in Qwen’s Omni stack.

Primary sources: Copilot computer use, Copilot model deprecation, Qwen3.8-Omni-Flash, and arXiv 2609.25611. Adjacent: Claude computer-use GA and agent eval sandbox checklist.

Copilot computer use

• Public preview in Copilot CLI and the GitHub Copilot app on macOS and Windows (October 1, 2026).
• Actions: read accessible content and visual context, click, type, scroll, drag, and cross-app GUI workflows.
• CLI: `/computer on`, `/computer show`, `/computer off`. App: Settings > Computer Use.
• Approval before controlling an app; always-allow lists can be reviewed or reset. macOS needs Accessibility and Screen Recording permissions.
• Organization-managed settings can disable the feature.

GitHub’s September VS Code notes (1.136–1.140) also add creating pull requests from Copilot, Claude, or Codex agent sessions and running agents in Dev Containers on SSH, Tunnel, and WSL. Computer use is the higher-risk control. PR-from-session is the higher-leverage workflow change for teams that already trust agent diffs.

October 2 Copilot model cutover

Enterprise admins may need to enable replacements in Copilot model policy before they appear in selectors. Automations that hard-code retired IDs will fail. Pair this cutover with the Sol / Opus 5.5 / Grok 4.7 routing board.

Qwen3.8-Omni-Flash

Alibaba launched Qwen3.8-Omni-Flash on September 20, 2026 as a native omnimodal agent: text, image, audio, and video in, 1M-token context, with the goal of planning, calling tools, and delivering finished AV work. Alibaba reports large gains versus Qwen3.5-Omni-Plus on audio-visual and agent benches, and a sharp drop in estimated hourly audio and audio-visual input cost versus prior Omni pricing.

• Agentic long-form video: coarse-to-fine evidence gathering instead of full-file prefills (Alibaba reports higher OmniVideoBench accuracy at ~46% fewer tokens in agentic vs static understanding).
• Meeting-to-action: speaker-aware minutes, then tool calls such as mail or coding from the same session.
• Production workflows: Music2MV, short-drama translation, long-form film commentary via Qwen-MM-Plugins.
• Open-source Qwen-Live-Harness for real-time omnimodal interaction (described in arXiv 2609.25611).

This is not a substitute for GPT-6.1 Sol or Opus 5.5 on repository work. It is a sub-agent for media-heavy loops, similar to how DeepSeek V4.1 Flash targets long text-agent KV cost.

What this means for developers

• Pilot Copilot computer use on a throwaway macOS/Windows VM with a non-admin user, not on a developer’s daily machine.
• Script `/computer off` in managed images until policy is signed.
• Replace retired Copilot model IDs in `.github`, editor settings, and Copilot CLI config.
• If evaluating Omni-Flash, use the Live Harness and MM-Plugins rather than stuffing hours of video into a text agent.

What this means for businesses

• GUI-only line-of-business apps (expense, ERP thick clients, desktop QA) become automatable without waiting for an MCP server. That is a productivity win and an audit problem.
• Disable computer use in Copilot org settings until DLP, screen-recording consent, and residual-data policy exist.
• AV localization and meeting-ops teams can trial Omni-Flash as a production worker with human review on outbound mail and published video.
• Expect Copilot spend to move toward Opus 5.5 and Gemini 3.8 Flash as old IDs disappear.

Pilot checklist

1. 01
   Finish Copilot ID cutover

   Map deprecated models to Gemini 3.8 Flash, Kimi K3, or Opus 5.5 and enable them in enterprise policy.
2. 02
   Decide computer-use posture

   Default deny org-wide. Allow only a named pilot team on isolated endpoints.
3. 03
   Write an allowlist of apps

   Expense, browser QA, and internal thick clients first. No password managers, banking, or prod admin consoles.
4. 04
   Log approvals

   Keep who granted always-allow, which apps, and when it was reset.
5. 05
   Separate Omni workers

   Route AV jobs to Qwen Omni-Flash or Gemini Live. Keep GitHub coding on Copilot/Claude/OpenAI text models.

Conclusion

Copilot computer use extends agents into software that will never grow an MCP server. Qwen3.8-Omni-Flash extends agents into hours-long audio and video. Both are production-shaped only with deny-by-default policy, model-ID hygiene, and a split between coding planners and media workers. Turn the features on after those three exist, not before.

Sources: github.blog ; github.blog ; github.blog ; alibabacloud.com ; arXiv paper