In the ten days between September 16 and September 26, 2026, three of the most consequential launches in the brief history of consumer AI happened within a week of each other. Microsoft unveiled a rebuilt Copilot “super app” with three new modules — Home, Code, and Autopilot — and told the world it intends to “rebuild Copilot as an entirely new work operating system.” DeepSeek shipped Harness Desktop in public preview for both Windows and macOS, with four modes (Standard, PTC, Minimal, Create) and a plugin philosophy the team summarized as “Everything is a Plugin.” Alibaba, at the 2026 Apsara Conference, unveiled Qwen Book, the first PC-class device marketed explicitly as an “AI agent computer,” and announced a collaboration with the Omarchy open-source project to build an agent-native desktop OS on top of it.
Three companies, three platforms, one convergent bet: the operating system is the product now, and the reasoning model that runs on it is the kernel. The DeepThink engine — already the most widely deployed open-reasoning stack on the planet — sits at the center of this transition. The question is what role it plays: kernel, runtime, or application layer.
The Three Launches in One View
| Date | Launch | Form Factor | Distribution | Core Agent Unit |
|---|---|---|---|---|
| Sept 16, 2026 | Microsoft Copilot Super App | Web + Windows + Office | Bundled with Windows 11 26H2 | Copilot Home, Copilot Code, Copilot Autopilot |
| Sept 23–25, 2026 | DeepSeek Harness Desktop | Native desktop app | Free preview, Win x64 + macOS arm64/x64 | Harness Standard / PTC / Minimal / Create modes |
| Sept 22–24, 2026 | Alibaba Qwen Book | Laptop + custom OS | China first, GCC roll-out Q1 2027 | Qwen Intelligence agent stack on Omarchy base |
The three announcements are not the same thing. Microsoft is rebuilding an existing productivity surface into an agent hub. DeepSeek is shipping a developer-native agent harness that treats every model, tool, and skill as a plugin. Alibaba is shipping hardware designed from the keyboard up to make agent invocation the primary user action. But the architectural overlap is striking: every one of them assumes that the user-facing agent is a persistent, stateful runtime, not a chat window.
Microsoft Copilot: The “Work Operating System” Play
Satya Nadella’s framing — “an entirely new work operating system” — is more than rhetoric. Copilot Home, the new entry point, integrates Chat and Cowork and pulls Word, Excel, PowerPoint, and Outlook directly into the agent’s tool surface. Copilot Code allows users to create applications, track tools, data dashboards, and automation flows through natural language. Copilot Autopilot is the most novel of the three: it runs while the user is not interacting, executing follow-up communications, managing workflows, scheduling meetings, and pushing projects forward without explicit prompts.
This is the autonomous-agent-on-the-desktop model. Microsoft is betting that the next version of Windows is a runtime for persistent agents, not a window manager for documents. The technical bet is that GPT-6 (running on Azure), Anthropic Claude (running on AWS via partnership), and the open-weight ecosystem (running on local NPUs via DirectML + ONNX) can share the same Copilot shell without losing coherence. The commercial bet is that enterprise IT departments will pay $99 per user per month for Copilot Autopilot alone, given the productivity it promises.
The DeepThink question is: where does DeepSeek fit? Microsoft’s stack currently routes through OpenAI for the default Copilot backend, with Anthropic as a secondary provider for coding workflows. The third slot — for open-weight reasoning on the local NPU — is still open. DeepSeek’s V4.1-Flash with its 8B/16B-active asymmetric architecture is the only open-weight model that currently meets the latency and quality bar for a persistent desktop agent. The integration is technically possible; the commercial question is whether Microsoft wants another foundation-model provider in the stack it is trying to monetize.
DeepSeek Harness: The “Everything is a Plugin” Play
Harness Desktop is the most architecturally opinionated of the three launches. The four modes are not separate products; they are different plugin compositions of the same underlying agent loop:
- Standard. Default mode for code, files, and document work. The agent calls retrieval, edit, and terminal tools as needed.
- PTC (Parallel Tool Calling). Emphasizes batch tool invocation with filtering, deduplication, aggregation, and summarization across results. Optimized for tasks that touch many tools at once.
- Minimal. Terminal-only. The agent uses only the terminal tool. Used as a baseline to test and compare underlying model behavior.
- Create. User-customizable mode. The agent can write plugins to add new functions or interfaces, combining tools and prompts into custom modes.
The “Everything is a Plugin” philosophy is the design principle. Models, tools, skills, sessions, sandboxes, storage, agent loops, scheduling, and UI are all exposed through a single plugin system called Cordis. The implication is structural: every part of Harness is replaceable. A user can swap in a different reasoning model, a different terminal, a different storage layer, or a different scheduling policy without rewriting the agent loop.
This is the philosophy that aligns with DeepThink’s open-trace-by-default design. If reasoning is an artifact, then the reasoning runtime should be composable, not monolithic. Cordis is the runtime-level manifestation of the same logic that produces exportable reasoning traces. The strategic implication: DeepSeek is not building an OS, it is building the plugin substrate that any agentic OS can compose around. Qwen Book runs on Omarchy, which runs on Linux, which runs on hardware — but every one of those layers can include Harness as a plugin layer.
Alibaba Qwen Book: The “Agent Computer” Play
Alibaba’s bet is the most aggressive: ship the hardware. Qwen Book is a laptop-class device with a Skill keyboard array, a global AI key, and a voice/stylus interface. The operating system is Omarchy, an open-source project for agent-native desktops, in collaboration with Alibaba Cloud. The default agent runtime is Qwen Intelligence, Alibaba’s full-stack mobile/desktop agent layer covering task planning, cross-application execution, and image creation.
Alibaba Cloud CEO Wu Yongming’s framing was direct: “Machine intelligence is becoming a commodity. In the future, the total amount of thinking supplied by machines will be more than 1,000 times the total amount of human thinking. Today, that number is less than 3%.” The Qwen Book launch is the productization of that thesis — a device whose primary mode of operation is agent invocation, not document editing.
For DeepThink, the most interesting move is the Omarchy partnership. Omarchy is an open-source, community-driven project; its roadmap is public, its governance is distributed, and its plugin model is permissive. By aligning with Omarchy rather than building a closed OS, Alibaba is positioning Qwen Intelligence as one of several possible agent runtimes that can sit on top of the hardware. DeepSeek Harness could, in principle, ship as an alternative runtime on Qwen Book hardware. The economics are interesting: Alibaba wants the device to win, not necessarily the agent stack. That openness is precisely the wedge DeepThink needs.
The Three Bets Are Not Mutually Exclusive
The most important structural fact about the current moment is that none of the three launches is the canonical agentic OS. The agent stack is fragmenting into layers:
| Layer | Microsoft | DeepSeek | Alibaba | Common Standard? |
|---|---|---|---|---|
| Hardware | OEM partner | None (runs on any) | Qwen Book | No |
| OS | Windows 11 26H2 | Win + macOS native | Omarchy (Linux) | No |
| Agent Runtime | Copilot shell | Harness + Cordis | Qwen Intelligence | Pluggable |
| Reasoning Model | GPT-6 / Claude / open-weight | V4.1-Flash + V4-Pro | Qwen3-Max + Step 5 | Composable |
| Tool Surface | Office + Graph API | Plugin marketplace | MCP + Skill ecosystem | Converging on MCP |
The bottom two rows are where convergence is happening. The Model Context Protocol (MCP), originally developed by Anthropic and now adopted across the industry, is becoming the standard wire format for tool invocation. The reasoning model layer is becoming composable: any agent runtime can call any reasoning model through an adapter. These two facts mean that the value is migrating up the stack from the reasoning model (commodity) and the OS (commodity) to the agent runtime and the workflow integration.
Where DeepThink Wins
Three places DeepThink wins in this transition:
- The reasoning kernel. V4.1-Flash’s 8B/16B-active design, with the upcoming V4.1-Pro scaling path, is the only open-weight stack that can run persistently on-device while serving frontier-class reasoning. Every agentic OS needs a default reasoning model; DeepSeek is the only open-weight candidate that meets the latency bar.
- The audit layer. As AGMAI-style reasoning governance spreads from mathematics into science, law, and medicine, the only structurally sound foundation is open reasoning traces. Harness exposes the trace by default; Copilot and Qwen Intelligence expose it only through enterprise SKUs. In regulated industries, that asymmetry compounds.
- The agent runtime substrate. Cordis’s plugin-everything architecture is more portable than Copilot’s monolithic shell and more production-ready than Qwen Intelligence’s still-emerging stack. It is the only major agent runtime that is explicitly designed to be embedded in other agentic OSes.
The Takeaway
The agentic OS race is not a winner-take-all market. It is a layered stack in which each layer has its own competitive dynamics. Microsoft is winning the enterprise default; DeepSeek is winning the developer-and-open-weight default; Alibaba is winning the China-region hardware default. The DeepThink engine sits in the middle, where the reasoning kernel matters most.
For builders, the practical implication is that no single agentic OS will dominate globally, but a small number of agent runtimes will sit above a heterogeneous OS and hardware base. Building on DeepSeek Harness — with its pluggable reasoning model, its open reasoning traces, and its cross-platform runtime — is the most defensible position. The agentic OS wars will be won by the agent runtimes, not by the operating systems. The OS is the kernel, the runtime is the agent, and the reasoning model is the substrate. DeepThink is the only stack that has been designed for all three layers from day one.