weekly.ai-news.cz
The AI week, distilled.
Week 41 · 2026
Microsoft AI | 5–11 October 2026

Microsoft expands agent workflows and adds controls for local AI

Microsoft announced new Copilot app-building and CRM capabilities alongside a Windows architecture combining local and cloud AI. GitHub made Copilot sandboxing generally available and added organizational billing controls for code review. Enterprise teams need to distinguish production-ready controls from preview features before widening deployment.

01

Copilot Studio combines apps, workflows, and agents

Microsoft expanded Copilot Studio to combine apps, workflows, and agents in one environment. Apps and GitHub Copilot harness hooks are in preview, while Foundry IQ integration and the agent-readiness Review panel are generally available.

  • Use a controlled pilot to assess whether the combined environment can replace separate interface, workflow, and agent tools.
  • Validate source permissions and data-access controls before connecting enterprise knowledge through Foundry IQ.
  • Keep preview apps and harness hooks outside production commitments until your team has assessed support and operational limitations.
Source — Microsoft Copilot blog Copilot StudioEnterprise workflows
02

GitHub Copilot local sandboxing reaches general availability

GitHub made local sandboxing generally available in Copilot CLI, the Copilot app, and VS Code sessions using Agent Host. Enterprise policies can enforce restrictions on agents' access to files, networks, credentials, and other system capabilities.

  • Require sandboxing before allowing coding agents to execute tasks on workstations containing proprietary code or credentials.
  • Test file and network restrictions against internal package registries, build services, and developer workflows.
  • Check policy enforcement across Windows, macOS, and Linux before expanding deployment across development teams.
Source — GitHub Changelog GitHub CopilotAgent security
03

Windows combines local and cloud AI with agent containment

Microsoft outlined a hybrid Windows AI architecture combining local models and actions with cloud processing. Microsoft Execution Containers became generally available with Agent 365 integration for agent isolation, identity, governance, and management.

  • Inventory eligible Copilot+ PCs before planning workloads that depend on local models or device context.
  • Define permitted file, network, and local-action access before enabling agents on managed endpoints.
  • Map which processing remains on the device and which reaches the cloud rather than treating hybrid AI as a data-residency guarantee.
Source — Microsoft Source WindowsAI governance
04

Microsoft adds CRM skills to Copilot

Microsoft announced 30 pre-built Copilot skills across Dynamics 365 Sales, Customer Service, and Customer Insights. Sales Development Agent is in public preview and is scheduled to begin consuming Copilot credits on 16 November 2026.

  • Test whether the skills cover your sales and service teams' actual Dynamics 365 processes before commissioning custom integrations.
  • Define approval and escalation points before allowing agents to act across CRM workflows.
  • Estimate Copilot-credit consumption for expected sales-development volumes before the scheduled November billing change.
Source — Microsoft Copilot blog Dynamics 365Copilot
05

GitHub centralizes Copilot code-review billing and access controls

GitHub added options to charge Copilot code reviews to an organizational cost center instead of individual member quotas. Administrators can also restrict review requests from users whose licenses sit outside the organization or enterprise.

  • Assign code-review spending to organizational cost centers to improve budgeting across repositories and teams.
  • Review contractor and subsidiary licensing arrangements before restricting externally licensed users.
  • Align review-request permissions with repository governance rather than relying on developers' individual subscriptions.
Source — GitHub Changelog GitHub CopilotCost management
06

Microsoft releases MAI-Code-1.1-Flash with a local execution option

Microsoft released MAI-Code-1.1-Flash for GitHub Copilot and local execution in quantized form. Microsoft reports comparable coding performance with 25% fewer tokens and recommends more than 120 GB of RAM for best local performance.

  • Benchmark representative internal coding tasks before using Microsoft's token-efficiency claim in a business case.
  • Compare the required workstation memory and operating costs with cloud inference before procuring local hardware.
  • Validate network behavior and surrounding development tools before treating local inference as a source-code confidentiality control.
Source — Microsoft Source Latin America Coding modelsLocal AI
07

Consumer Copilot introduces Home, Code, and Autopilot

Microsoft announced a redesigned consumer Copilot organized around Home, Code, and Autopilot. Autopilot adds persistent delegated work, including topic monitoring, project continuation, and recurring tasks.

  • Do not assume the consumer announcement establishes availability or licensing for your Microsoft 365 tenant.
  • Update employee guidance on entering company data into consumer agents that retain context or perform recurring work.
  • Assess sharing, data connections, and Czech-language behavior before allowing employee-created apps into business workflows.
Source — Microsoft Source CopilotDelegated agents
08

Claude Haiku 5.5 becomes available in GitHub Copilot

GitHub added Anthropic's Claude Haiku 5.5 to Copilot, including Business and Enterprise plans. Supported clients include VS Code, Visual Studio, Copilot CLI, and the Copilot cloud agent.

  • Evaluate the model on frequent, smaller tasks such as quick edits and terminal work rather than assuming suitability for complex changes.
  • Review model policies before making another provider's model available to developers handling sensitive repositories.
  • Measure task quality and usage-based charges together before setting team-wide model defaults.
Source — GitHub Changelog GitHub CopilotModel choice
09

GitHub launches ReviewBench for AI code-review evaluation

GitHub introduced ReviewBench, an open benchmark for AI code-review agents. It covers 219 pull requests from 187 repositories across 19 programming languages.

  • Use the published methodology as one input when comparing AI code-review products for procurement.
  • Check how closely the benchmark's languages and repositories match your enterprise application portfolio.
  • Supplement benchmark results with internal pull requests to measure missed defects and unnecessary review comments.
Source — GitHub Blog AI evaluationCode review
10

GitHub's weekly release adds parallel agents and local-model discovery

GitHub's weekly Copilot release added side-by-side agent sessions in VS Code and local-model discovery through Ollama in Copilot CLI. It also added worktree cleanup tools and separate accounts for Copilot licensing and repository access.

  • Test parallel agent sessions with branch and worktree rules that prevent conflicting changes.
  • Evaluate locally hosted models under existing endpoint-security policies rather than assuming local discovery guarantees data containment.
  • Check whether separate licensing and repository accounts resolve access arrangements for contractors or subsidiaries.
Source — GitHub Changelog Developer workflowsLocal models
11

Microsoft France opens its fourth GenAI Studio cohort

Microsoft France opened applications for a fourth GenAI Studio cohort of 15 startups. Participants receive Azure credits and access to Azure AI Foundry's model catalogue, including Mistral models.

  • Treat the accelerator as a France-specific program rather than an offer of Azure credits to Czech enterprises.
  • Assess Mistral models against Czech-language business tasks if your Azure strategy includes multiple model providers.
  • Verify hosting region and contractual data handling separately from a model developer's European location.
Source — Microsoft Source France Azure AI FoundryEuropean ecosystem

Copilot changes this week

From Message Center and the public Microsoft 365 roadmap — admin actions first.

All 49 Copilot changes on ai-news.cz →

Microsoft Learn: what changed

The most significant documentation changes this week.

All 361 documentation changes →