Microsoft's Hybrid-Intelligence Copilot: Local Context, Local Actions, Local Models — and an Agent-Containment Gate (MXC)

Microsoft's Hybrid-Intelligence Copilot: Local Context, Local Actions, Local Models — and an Agent-Containment Gate (MXC)

On October 7, 2026, Microsoft announced that Copilot on Copilot+ PCs gains three local capabilities under the banner of "hybrid intelligence" (Pavan Davuluri, EVP Windows + Devices, via the Windows Experience Blog; reported by Unite.AI, fetched this cycle):

• Local context: "With local context, and with the user's permission, Copilot can understand relevant content on the PC, including files and recent activity."
• Local actions: "With local actions, Copilot can work on the user's behalf across the operating system, including organizing files, assessing device diagnostics, troubleshooting issues, coding, and carrying out workflows directly on the device."
• Local models: Copilot "can tap AI models running on the PC, using local compute for the right tasks while combining it with cloud intelligence when needed."

Model specifics: MAI Code 1.1 Flash (137B total / 6.8B active) runs on device at 3-bit with a 256K context window; an upcoming NVIDIA Nemotron model >70B quantized to 2-bit (~20 GB); DeepSeek V4 Flash (284B) on RTX Spark hardware. GitHub's HydraFusion router is being extended to tap local models (experimental preview later in October in the GitHub Copilot app, CLI, and VS Code). Microsoft framed the push around cost — "customers' needs are outpacing what their cloud budgets can support" — claiming >2 trillion local inferences/month across Copilot+ PCs and >40% of business laptops being built as Copilot+ PCs.

Microsoft Execution Containers (MXC) went GA the same day — a policy-driven containment layer for agents: developers declare the files and network destinations a workload requires via a JSON schema, and "the containment layer keeps policy outside the agent workload's control, so the agent or its generated code cannot grant itself additional access." Agents already supporting MXC: OpenAI's Codex, GitHub Copilot, OpenClaw, Replit, LM Studio, NVIDIA's OpenShell, Unsloth AI. Planned support: Anthropic Claude Code, Perplexity, Manus, and Meta's Muse (coming as a native Windows app with MXC integration).

Hardware trajectory: the Surface Laptop Ultra (RTX Spark, up to 128 GB unified memory, models exceeding 120B parameters locally) ships October 16, 2026; DGX Station desksides that "can run frontier-class models at more than a trillion parameters locally" arrive on Windows later in 2026.

What it means for the map: the local-first, model-agnostic layer Hey, Lefty occupies (Market Map & Positioning: Where Does Hey, Lefty Fit?) is now being compressed from the OS side, not just the host-app side (cf. the existing Perplexity Portable Computer watch). Microsoft ships native local context + OS-wide actions + on-device models to the Windows installed base, and inserts MXC as the containment gate third-party agents must pass to act on Windows — with Perplexity and Muse already queued as planned supporters. Two positioning consequences: (1) "runs locally on your machine" stops being a differentiator against the default OS itself; the durable differentiators are model-agnosticism (Microsoft routes to its own MAI/Nemotron/DeepSeek stack via HydraFusion) and cross-harness portability; (2) MXC compliance becomes a distribution requirement for any agent acting on Windows¹ — a new compliance surface alongside the CFAA/agent-identification regimes tracked in Agentic Browsing and the CFAA: The Ninth Circuit's Amazon v. Perplexity Ruling — and the Post-Legal Blockade of Agents. EEA availability is unstated ("availability and timing varying by device, market") — registered as a watch.