Microsoft AI Update October 2026

Microsoft AI Update — October 2026

John Savill reviews the month’s AI industry themes and Microsoft-related product, model, and infrastructure updates.

Industry themes

• Scaling AI responsibly: A central challenge is expanding AI use while maintaining accountability, reliability, governance, observability, and transparent costs. Observability supports auditing and improvement—for example, deciding whether to change prompts, retrieval-augmented generation (RAG), tools, memory, or models.
• Redesigning work: The presenter cites research indicating that many leading AI users are redesigning workflows rather than simply adding AI to existing processes. The goal is to rethink work around the combined strengths of people and AI, and to reconsider roles as agents become part of the workforce.
• Managing operating costs: As the number of models and the complexity of tasks grow, organizations need to match the model, reasoning level, and workflow to the job to get the desired result efficiently.
• Sovereignty and resilience: These are growing considerations for organizations choosing AI systems and providers.
• Trust through controls: The presenter emphasizes building systems that can safely manage models through governance, security, isolation, and monitoring, rather than relying on trust in a model’s behavior.

Videos released this month

Savill highlights videos covering:

• A short introduction to MCP.
• A review of the Copilot app and its combined chat, collaboration, coding, and automation features.
• Why agents need computer-use capabilities when APIs or MCP servers are unavailable.
• Agent control planes and why they matter.
• Reasons agents fail.
• Using AI to accelerate security defense.

Models and AI infrastructure

• MAI Code 1.1 Local: A quantized coding model intended to reduce memory requirements. In supported GitHub Copilot scenarios, it can run locally on NVIDIA RTX Spark hardware. Developers can select it directly or let Copilot route work between local and cloud models based on the task and context.
• MAI Base 2: Microsoft’s mixture-of-experts foundation model, described as having 2.8 trillion total parameters, with about 80 billion active for a given task. Microsoft plans to use it as a foundation for later models.
• Decision-making tools: The video contrasts models designed to produce judgments or probabilities with OpenAI’s decision-making API, which provides similar decision capabilities through an API built on an existing model.
• GPT-6.1 Terra: Available through Foundry and GitHub Copilot, with improvements in visual reasoning and computer interaction. The presenter also notes writing improvements, a one-million-token context window, text and image input, and adjustable reasoning levels.
• MAI Image 2.6 and Flash: The standard model is positioned for precise, complex image editing and high-quality text rendering. Flash is described as faster, smaller, and less expensive, making it suitable for iterative, interactive use.
• Mistral Large 4: A large mixture-of-experts model with open weights, developed in Europe—a potential consideration for organizations with sovereignty requirements.
• Sonnet 5.5: Highlighted for creating polished PowerPoint presentations through Copilot, including working with slide layouts and templates.
• Maia 300: Microsoft’s next-generation AI inference chip. The stated aim is to make it a primary inference platform from 2027 onward; the existing Maia 200 is already used in some Microsoft services.

Foundry, agents, and Copilot Studio

• Microsoft Copilot harness in Foundry: Developers can use the harness—the component that manages an agent’s cycle, tools, skills, and memory—to build agents with capabilities similar to those used in Microsoft’s Copilot products. Microsoft is standardizing the toolkit across GitHub Copilot, Copilot Code, and Copilot Studio.
• Browser automation: Now generally available, it lets agents interact with websites through a managed browser when a suitable API or MCP server does not exist.
• Multi-LoRA adapters: Multiple lightweight, specialized adaptations can be applied to the same base model and managed compute instance. This can make it cheaper and easier to serve models customized for different behaviors or terminology.
• SIP for voice agents: Foundry-hosted voice agents can connect to customers’ own SIP telephony services.
• Customer conversation capabilities in Copilot Studio: The conversation harness supports personalized customer-service agents across channels. Specialized agents can handle tasks such as discounts, orders, or service requests, work together, and hand off to a person when needed.
• Agent identities: Each Copilot Studio agent can have its own identity, enabling least-privilege access and better auditing of the data and tools it uses. Agents can also be invoked in Microsoft 365 apps using an @ mention.
• Business-process memory: Agents can retain user preferences and process context between interactions.
• Triggers: Agents can run in response to events, such as an incoming email, or on a schedule. The presenter notes that usage is charged for the work performed.

Microsoft 365 Copilot

• Copilot Projects: Organize project-specific instructions, chats, files, and contacts.
• Copilot Workspaces: Extend notebook-style collaboration so people and AI can work together. Word, Excel, and PowerPoint files can be edited in the workspace, while other file types can be linked. Workspaces can also produce audio overviews.
• Role-based skills and plugins: Co-work offers skills for areas such as sales, services, education, law, and finance. Examples mentioned include accounting and data-cleaning capabilities in Excel, a legal plugin for Word, and a Life Sciences connector.
• Legal and Finance offerings: Legal Pack adds specialized legal skills and partner plugins for drafting and reviewing documents. Copilot for Finance adds skills and plugins aimed at accounting, capital markets, and financial reporting.
• Referencing earlier conversations: Users can include previous Co-work conversations as context for new work.
• Teams meeting recaps: According to the presenter, AI can provide a summary and identify tasks and owners without saving a transcript.
• Work IQ MCP retrieval: A new “get” option returns relevant source information from Microsoft 365—such as emails, files, meetings, Teams messages, and contacts—in a format intended for generative models, with citations and metadata. Access is delegated and limited to what the user can access.
• M365 Insights agent: A natural-language agent for exploring organizational agent usage, including adoption, engagement, retention, and comparisons between agents.

GitHub Copilot

• New GitHub Copilot app: Presented as a broader workspace, not just a code-generation tool; it can also help with specifications and telemetry analysis.
• Copilot Fusion: Routes and orchestrates work across models to balance quality, speed, and cost. It can assign drafting and review to different models, use feedback loops, escalate to more capable models when needed, and use local inference where supported.
• Code Review Max: Offers a deeper, higher-quality review option for important tasks.
• Copilot Projects: Supports work across multiple repositories and provides a central place to track parallel sessions.
• Session continuity: A cloud-running agent session can be continued across devices and environments, such as moving from a phone to VS Code.
• Copilot Automations: Cloud-based, event- or schedule-triggered workflows can handle tasks such as code review or issue sorting. They can suggest improvements based on past performance, with a human required to review and approve them.
• Infrastructure changes: GitHub has adapted its architecture and added hardware to handle the higher volume and different patterns of agent-generated pull and merge requests.

Agent governance, APIs, and security

• Agent 365 and API Management: The preview integration connects Agent 365’s agent oversight capabilities with Azure API Management (APIM). It can discover relevant API Management and AI Gateway resources, expose them through an inventory, and apply policies based on instructions.
• APIM MCP and AI Gateway features: APIM adds MCP inventory, controls, observability, and threat protection. The presenter distinguishes the general V2 tier from the AI Gateway tier, which provides more AI-specific and tool-level controls.
• Project Perception: The security initiative is expanding to address risks from internal agents, including unexpected or manipulated behavior. Signals from Project Perception and Agent 365 can inform changing risk assessments and policies, dynamically restricting an agent’s access or blocking it when behavior deviates from expectations.

Main speaker and sources

• Main speaker: John Savill, the video’s presenter.
• Sources and organizations referenced: Gartner, McKinsey, Accenture, Microsoft, OpenAI, GitHub, Mistral, and other AI vendors.