Today, Microsoft Foundry Managed Fine-Tuning expands its collection with four new open-source models and brings MAI-Code-1.1-Flash to Foundry resources, all in public preview. Alongside them,Multi-LoRA Managed Compute deployments enter public preview, letting compatible adapters share base-model weights and GPU capacity instead of each needing a separate serving allocation.
More open-source choice for your next custom model
Muse-Glimmer-30B, qwen3.8-27b, qwen3.6-35b-a3b and gpt-oss-120b join Foundry's managed open-source model collection, with supervised fine-tuning (SFT) support across all four. You can now bring your example-driven customization workflow to more backbones, choosing for tool-use reliability, coding efficiency, extended planning or reasoning capacity. Here's what each option brings:
Model
Why put it on your shortlist?
Muse-Glimmer-30B
Resilient tool orchestration: precise tool calls, interpreting unexpected results and recovering when a workflow goes off-script.
qwen3.8-27b
Extended research and planning: a dense backbone for sustained autonomous work that adapts to feedback from its environment.
qwen3.6-35b-a3b
Repository and frontend coding: a sparse mixture-of-experts backbone for navigating codebases and making coordinated changes.
gpt-oss-120b
Reasoning headroom: a higher-capacity alternative to the existing gpt-oss-20b option for difficult analytical and problem-solving tasks.
Use SFT to tailor these models to your application's needs:
Coding conventions: organization-specific library usage and preferred implementation patterns.
Structured outputs: JSON schemas and extraction responses used by downstream applications.
Domain workflows: product terminology, support tone, and classification or triage decisions.
Distillation: use teacher-generated example responses to specialize a smaller model for your task.
All four also support reinforcement fine-tuning (RFT) if your workflow calls for more advanced, reward-based optimization.
MAI makes its fine-tuning debut in Foundry
In addition to the open-source models above, we're introducing MAI-Code-1.1-Flash, the first Microsoft model available for Managed Fine-Tuning on Foundry resources. With the base model already available in GitHub Copilot, this compact, coding-focused model brings particular strengths in software development tasks. Its emphasis on responsiveness and token efficiency makes it a compelling backbone for repeated edit-build-test loops, where inspecting code, proposing edits and responding to compiler or test feedback all contribute to the task's latency and token budget.
MAI-Code-1.1-Flash supports RFT only in this preview. Use it to target outcomes you can verify: a change that passes task-specific tests, correctly uses an internal software library, or satisfies your repository's acceptance checks. Build rewards around correctness and the required scope of the change, rather than whether an answer reproduces a reference patch. That leaves room for different valid solutions while keeping the customization anchored to your engineering requirements.
Share hardware across adapters to reduce serving costs
Previously, Foundry's managed fine-tuning outputs went to serverless pay-as-you-go deployments: input/output token charges plus hosting fees for each fine-tuned deployment. More traffic increased usage charges; more deployed variants added hosting costs. With Multi-LoRA adapter deployment on Managed Compute, compatible adapters share LoRA-enabled base-model deployment and GPU capacity. You pay an hourly rate for that hardware, rather than per-token inference charges, so multiple customizations can use one serving allocation instead of funding separate GPU footprints.
Three task-specific qwen3.8-27b adapters, for example, can share a LoRA-enabled base deployment while remaining separately addressable. When their combined traffic fits, adding an adapter doesn't mean another hardware bill. You avoid duplicating base-model serving capacity and can improve utilization by pooling requests across variants, instead of paying separate serverless hosting fees for each one.
The tradeoff is utilization: running hardware is billed even when idle, so size for aggregate concurrency and context lengths. More traffic can require additional capacity, while serverless remains an option for intermittent workloads on eligible models. The managed compute deployment path supports Muse-Glimmer-30B, both new Qwen models and gpt-oss-120b. Adapters must match the same base model and version; the base deployment needs a compatible LoRA-enabled template. Accelerator quota is required.
Put the new models to work in your agents
The expanded model collection helps you pick a backbone suited to your tasks; Multi-LoRA lets compatible customizations share serving capacity. These choices come together in agent applications, where different roles can need different model behavior. A coding assistant and a code-review agent, for example, can use separate adapters on the same supported OSS base model while sharing the hardware that serves them.
Fine-tuning customizes model behavior, but the goal is to improve the agent application that uses it. Connect your tuned models to your agent applications in Foundry and evaluate them alongside the instructions, tools and workflows they'll use in production. That keeps customization focused on better application outcomes, not just better training results.
Over the past few weeks, I’ve seen a massive push from developers to adapt their apps for iPhone Duo ahead of its launch date.
I’ve been especially impressed by the indie dev and Build in Public communities, with many developers going well beyond what’s required for basic Duo support.
There was also some interesting cross-pollination between iPad and Duo. Some developers capitalised on their recent efforts to add iPad support, extending it to Duo.
Others went the opposite way and added iPad support alongside their Duo updates.
But it’s not only the updates that caught my eye.
There are so many interesting Duo-inspired ideas and apps floating around, like this retro-style Contacts index or this notebook app.
Deck-man, the Sony Walkman app we mentioned in a recent issue, has gone proper viral, with big YouTubers featuring it in their reviews.
It’s clear that skeuomorphism is making a strong comeback with Duo! Conveniently, someone built FoldHunt to capture these Duo-inspired ideas and track which existing apps have been adapted.
This makes me feel like we’re back in the early days of the App Store, when we’d check in daily, eager to see what was new on the shelves. Hopefully, this feeling lasts for a while…
In fact, we might just have plenty of time to enjoy the anticipation.
iPhone Duo starts shipping today, but significant supply delays mean that many developers won’t get their devices any time soon.
Apple provided tech bloggers and YouTubers with review units ahead of time, which raises the question of whether a similar accommodation could be made for those developers who worked hard to provide Day One support.
Perhaps Apple could allow developers to pre-order a device through their Apple Developer accounts, with some stock reserved for this purpose?
I admit, I’ve no idea how the logistics of this would work, but there are many folks at Apple who are much smarter than me and could work something out.
Granted, Apple is running multiple Optimize your app for iPhone Duo in-person workshops at Developer Centers around the world, where devs can get hands-on time with the devices.
Keep your GitHub Actions workflow and give it an M5 Pro runner. Connect your repo to WarpBuild, then set runs-on: warp-macos-latest-arm64-6x in your macOS job and see the immediate improvement to build times and costs! Try it with $50 in free credits using code iosdev2026.
Sean Allen, a prominent iOS educator and YouTuber, announced that he’s shutting down his online school,
putting some courses on sale and making others free to download.
Sean explains: ‘iOS dev course sales plummeted due to AI’.
I don’t know what the long-term effect of AI on online education would be, but losing the human touch that educators like Sean brought, could be devastating indeed!
I hope Sean finds other ways to share his knowledge that are sustainable in this era of AI.
Nikhil Nigade shares his design explorations adapting his app Pockity for iPhone Duo.
This might be the first time I’ve seen someone use mesh gradients creatively to tie together a multi-screen app design.
I think it’s a clever approach that also looks fantastic!
Most of the iPhone Duo articles I’ve seen focus on adapting iPhone-first apps.
This article from Karolina Samorek stands out because it describes the experience of adapting an iPad-first app Lake (which, btw, is beautiful!).
Starting with an iPad layout doesn’t automatically mean it’s iPhone Duo-ready. But it’s reassuring to see that the lift isn’t too heavy.
I normally just skim through How-to articles unless I’m looking for something specific.
But this article by Sarun I read very carefully, especially with the current push to adapt apps for iPhone Duo.
Figuring out how to correctly set layout margins was one of the first things our team had to do in preparation for Duo, and believe me, it’s not as straightforward as it seems.
Florian built exploreswiftui.com to visually showcase the native capabilities of SwiftUI ‘in its cleanest form’.
It even has a dedicated page with iPhone Duo examples!
Resources like this play an important role in the Apple Developer ecosystem because they fill the gaps in the official documentation (of which there are many!). Cheers, Florian!
Did I ever mention that I love Mac-assed apps? 🙂
ImageFlow is a strong addition to the family of unapologetically native macOS apps.
And it’s open source!
If you aren’t familiar with node-based editing, I highly recommend giving ImageFlow a try!
I first learned about this concept when I tried Nuke a few years ago, and I remember thinking that it felt like visual functional programming.
Great job, Daniel Fortes!
There’s a new harness in Copilot Studio! The Customer Conversation harness is built for agents that talk directly with customers – over voice and chat, in the same agent – combining predictable, controlled conversation paths with generative AI, sub-agents and built-in escalation to people.
It arrived together with Microsoft Contact Center Intelligence (CCI), announced on October 21, 2026. Most of the headlines are about the contact center, but the bigger story for makers is the harness underneath it. CCI is the application that puts this harness to work in contact centers – and the same builder is available directly in Copilot Studio.
The harness is in public preview and rolling out now.
Let’s take a closer look at what the new harness brings, how it fits next to the other harnesses, where Contact Center Intelligence comes in, and how it differs from Teams Phone Agent.
In this post:
The Customer Conversation harness – what is new
The Customer Conversation harness is the runtime for customer-facing agents in Copilot Studio. An agent built with it is called a Customer Assist Agent.
It is purpose-built for customer-facing conversations: natural, low-latency, over voice and chat. And the same builder is available both in Copilot Studio and in the Contact Center Intelligence portal – the authoring experience is identical no matter where you start.
What makes this new:
One agent for voice and chat. Every agent supports both Speech and Chat by default. You don’t need to choose the modality up front – and you can give separate instructions for speech and chat.
Deterministic and generative together.Topic skills bring back the topics we know from the standard harness, so critical customer journeys (payments, identity checks, compliance steps) follow a fixed path. Generative skills let AI handle the more open-ended conversation. This combination is the key thing here.
Multi-agent by design. A primary agent greets the customer, classifies intent and routes to specialist sub-agents – billing, technical support and so on – and each has its own escalation path to a human.
Prebuilt contact center tools.Escalate (to Teams Phone, email or a live agent through a contact center) and Biometric Authentication (biometric authentication needs an environment with Dynamics 365 apps enabled)
Plenty of tool options. Connectors, MCP servers, REST APIs, AI prompts, workflows and connected agents (currently Microsoft Foundry agents). Skills and memory are supported.
Memory for returning customers. The agent can carry approved context from earlier conversations, so a returning customer doesn’t start from zero. Memory is structured, short-term (30-day retention), and with a shared contact identifier, context from a phone call is available in the next chat – and the other way around.
Build, test, deploy and monitor in one place. The agent has Build, Preview, Evaluate, Deploy and Monitor tabs. When you publish, Copilot Studio automatically creates a demo website for testing.
Deployment channels range from the demo website, web chat and Teams Phone to Dynamics 365 Contact Center, Genesys or your own contact center platform. So you can start with a website or a phone number – no contact center needed.
Four harnesses in Copilot Studio – which one is which?
Copilot Studio now uses the word harness for the runtime behind every agent. It sits between what you design and the AI model: it decides when to call the model, what to send to it, and which tools to call. The harness affects what the agent can do, where you can publish it, and how it is billed.
With the Customer Conversation harness, there are now four:
Standard
Copilot Chat
GitHub Copilot
Customer Conversation (preview)
Best for
Rule-based agents and structured conversations
Extending Microsoft Copilot Chat with company knowledge
Complex, multi-step business processes
Customer service over voice and chat
How it works
Follows the topics and paths you build
Answers from your knowledge inside Copilot Chat
Reasons toward a goal step by step and adjusts when a step fails
Primary agent routes to specialist sub-agents; fixed topics and AI skills together
Audience
Employees or external customers
Employees
Employees or external customers
Website chat or a phone number
Billing
Copilot Studio licensing for standard harness
Consumption, or included in Microsoft Copilot licenses
Copilot Credits
Copilot Credits
Copilot Chat harness might be a new term for many – it was for me too! The name comes from Microsoft’s own documentation: Extend Microsoft Copilot with the Copilot chat harness. In practice it is the agent for Microsoft 365 Copilot (declarative agent) that you build in Copilot Studio by selecting Microsoft Copilot on the Agents page.
Customer Conversation harness takes the predictability of the standard harness and adds the generative, multi-agent capabilities on top – built for talking with customers instead of employees.
A simple way to choose:
An employee agent inside Copilot → Copilot Chat
Every path should be fixed and predictable → Standard
The agent needs to carry out a whole business process → GitHub Copilot
The agent talks directly with your customers → Customer Conversation
Contact Center Intelligence – the harness at work in the contact center
If the harness is the engine, Microsoft Contact Center Intelligence is the application built around it for contact centers. Customer Assist Agents are built with the Customer Conversation harness, and CCI brings them – together with quality, insights and governance – on top of contact center platforms that Microsoft does not run. The idea is simple: keep the contact center platform you already have – telephony, routing, phone numbers, agent desktop – and add Microsoft’s agents, quality scoring and analytics on top of it. No rip-and-replace migration.
Image: Microsoft
Around the agents, the initial CCI preview includes:
Contact center integrations – connect supported CCaaS platforms with native connectors, APIs and event streams. Initial guidance covers Genesys Cloud plus a universal connector for other platforms.
Quality and operational insights – Quality Assurance evaluation of interactions, real-time monitoring and historical dashboards. Quality scoring can evaluate up to 100% of conversations instead of a sample.
Biometric authentication – voice biometrics to identify and authenticate callers, including consent and authentication data management.
Microsoft platform integration – reuse agents built in Copilot Studio and extend them with Microsoft Foundry.
The CCI portal is also where an agent is connected to a live contact center: deploying to a contact center is done there by an Omnichannel Admin, along with biometric authentication setup and quality, insights and governance management.
Teams Phone Agent vs Customer Conversation agent
Both answer phone calls with AI, so they can easily look like the same thing. They are not – they do different jobs, and different people set them up.
Teams Phone Agent is the AI receptionist. It is part of Teams Phone, and a Teams admin configures it in the Teams admin center. It takes the old “press 1 for sales” auto attendant menu further with a natural conversation: it screens calls, answers common questions, manages appointments and routes the caller to the right person or call queue. Its main job is to get the call to the right place.
The Customer Conversation agent solves the case. It is built in Copilot Studio by a maker, and it is designed to handle the customer’s issue from start to finish: verify the caller, look up their orders or invoices, follow fixed steps where the process requires them, use AI for open-ended questions, and hand over to a person when needed. The same agent works in web chat and on contact center platforms too.
Teams Phone Agent
Customer Conversation agent
Role
Receptionist: answers, screens and routes calls
Service agent: solves the customer’s issue
Who builds it
Teams admin
Copilot Studio maker
Channels
Phone
Voice and chat in one agent
Depth
Built-in Q&A and appointments
Sub-agents, fixed topics, AI skills, connectors, MCP, workflows
Caller verification
–
Voice biometrics and one-time codes against customer records
Handover to people
Routes to people and call queues
Escalates to Teams Phone, contact center or live agent with context
One way to think about it: the Teams Phone Agent picks up the phone; the Customer Conversation agent solves the case. An organization on Teams Phone could use Teams Phone Agent for general reception, and route billing, order or support calls to a Customer Conversation agent.
No sign-in on voice – does it mean just Q&A?
The documentation says that speech channels don’t support end-user sign-in, so knowledge sources that need the user’s own credentials work only in chat. That raises a fair question: is a voice agent then just a Q&A bot without access to real customer data?
No. It only means the caller can’t sign in to the call the way a web user signs in to a portal. The agent can still identify the caller and work with real data:
The agent verifies the caller. With Voice authentication turned on, the agent uses biometric authentication, authentication policies and one-time code challenges. You choose what identifies the caller – phone number, account number or another identifier – and match it against customer records in Dataverse (Contact and Account tables by default). Just like a human agent asking for your customer number and a code.
Tools fetch the data. Once the caller is verified, connectors, MCP servers, REST APIs and workflows look up that customer’s orders, invoices or cases – using the agent’s own connection, not the caller’s sign-in. Topic skills are perfect for this kind of fixed, controlled step.
The contact center passes context. When the agent runs on a contact center platform, the caller’s number and customer record are already known, and a verified identity carries over when the call goes to a human representative.
Chat can use real sign-in. The agent’s authentication mode controls how users sign in across the whole orchestration, so knowledge that needs the user’s own credentials still works in chat.
What does not work over voice is answering based on the caller’s own permissions – “show me the documents I can access.” That needs real sign-in, so it’s chat only.
So, on voice, the agent works like a human contact center agent: verify who is calling, then look up their data with the agent’s own access. This needs careful design – give the agent only the access it really needs, and verify before revealing anything personal.
Note: voice authentication requires an environment with Dynamics 365 apps enabled, so a plain Copilot Studio trial environment won’t be enough to demo caller verification.
B2C, B2B and inside the company
B2C
Utilities and telecom – outage reporting, invoice questions and contract changes. Payment arrangements follow a fixed path, general questions get AI answers.
Insurance – first notice of a claim over the phone, with voice biometrics so the caller doesn’t need to answer security questions again when transferred to a person.
Retail and e-commerce – order status, returns and delivery changes, with sub-agents for orders, returns and products.
Healthcare and public services – appointment booking, prescription refills, opening hours and guidance.
B2B
Product support – technical support for business customers
Distributor and partner desk – stock levels, order status and price checks through ERP connectors.
Field service – technicians calling from a customer site to look up spare parts or open a service order, hands-free.
Inside the company
HR service desk – benefits, enrollment, payroll questions and absence reporting
IT help desk by phone – password and MFA issues, outage status, and escalation through Teams Phone to second-line support.
Facilities and workplace – bookings and fault reports, even from a meeting room phone.
An internal service desk can also be a smart first project: no CCaaS integration needed, no external customer data, and a scenario almost every organization recognizes.
How building an agent works
Based on Microsoft’s documentation, this is what building a Customer Conversation agent looks like.
What you need
Copilot Studio and an environment where you can create agents.
Copilot Credits. The harness uses usage-based billing – building, testing and evaluating agents can consume Copilot Credits.
Optional: an environment with Dynamics 365 apps enabled for voice authentication, and Teams Phone licensing with a phone number for the Teams Phone channel.
Not needed to get started: a contact center. Connecting an agent to a contact center is done later from the Contact Center Intelligence portal, by an Omnichannel Admin.
The steps
Create – on the Copilot Studio Home page, under Other ways to build, select the Customer Conversation tile. Or go to Agents → New agent and pick the agent with the Customer Conversation badge.
Describe – tell Copilot Studio in natural language what the agent should handle. It generates a starting primary agent with a name, description and instructions, plus suggested knowledge, tools and channels. You can refine it over several messages, or start from a blank agent.
Instruct – check the description (max 1,024 characters, used for routing) and instructions (max 8,000 characters). The Speech and Chat tabs let you give separate guidance per modality – for example shorter answers and no lists over voice.
Build – add knowledge, Generative and Topic skills, tools such as Escalate, events (Conversation Start, End of Conversation, On Error), and sub-agents or connected agents for focused intents like billing.
Test – try conversations on the Preview tab and run test sets with pass/fail results on the Evaluate tab.
Deploy and monitor – publish on the Deploy tab. Copilot Studio creates a demo website automatically, so you can try the agent like a customer would. Then add channels like web chat or Teams Phone, and follow activity on the Monitor tab
Tip for the first test: use a public website as the knowledge source. Knowledge that needs the user’s own sign-in only works in chat, while a public site works in both chat and voice.
Things to know before a pilot
It is a preview. Documentation is prerelease and subject to change, and preview features aren’t meant for production use.
Credits are consumed while building too. Building, testing and evaluating all use Copilot Credits, not only production conversations.
There can be two bills. Some contact center vendors charge for API access or voice-channel licenses, and those costs are not part of Microsoft’s pricing. A pilot budget should include both.
Voice biometrics is biometric data. Bring in legal and privacy early – consent, retention and regional rules matter.
Design the data access carefully. Over voice, the agent fetches customer data with its own access, so least privilege and verify-before-reveal are a must.
Web grounding leaves the compliance boundary. The “Use information from the web” setting uses Grounding with Bing Search, which can send data outside your organization’s compliance boundary.
Final thoughts
Building a good customer-facing voice and chat agent has needed many separate pieces. The Customer Conversation harness puts them together in Copilot Studio: voice and chat in one agent, fixed paths where the business needs control, AI where the conversation needs flexibility, memory for returning customers, and a person available when needed.
Contact Center Intelligence is the first big application built on it – and a strong one. But the harness itself is the part to watch. It works with a simple website or a Teams phone number, on top of the contact center an organization already has, and as an internal service desk. That makes it much easier to start small.
This is great example how to build AI empowered solutions helping in business scenarios!
Hello Windows Insiders,
Today, we’re releasing new Windows 11 Insider Preview builds to the Beta and Experimental channels.
New builds this week
Select your channel below to view its release notes:
We've added the ability to filter Feature flags by enablement state.
[caption id="attachment_179240" align="aligncenter" width="1686"] Feature flags filtered by enablement state on the Windows Insider Program settings page.[/caption]
[Quick Settings]
Release channel: ExperimentalAccess your phone controls from Quick Settings
We're beginning to roll out a new Mobile devices experience in Quick Settings, making it easier to take quick phone actions (such as ring your phone) without breaking your flow on Windows.
With your supported Android phone connected, everyday actions are now just a click away. You can turn Do Not Disturb on before a meeting, change your phone's ringer mode, and play a sound to find your phone nearby.
It's a simpler, faster way to manage your phone from the place you already use for quick Windows actions, helping your PC and phone work better together throughout the day.
Capabilities can vary by phone platform and device configuration. The initial experience provides the most complete set of controls for Android phones, while iPhone support starts with a more limited status-focused experience.
[caption id="attachment_179242" align="aligncenter" width="534"] UI showing Mobile devices controls in Quick Settings for managing your connected Android phone.[/caption]
Check your internet speed directly from Quick Settings
We're making it faster and easier to check the performance of your internet connection in Windows. Windows Insiders can now run an internet speed test and see the results directly in Quick Settings, without being redirected to a browser.
Start a test from the new Speed test Quick Setting, from the Wi-Fi or Cellular flyout, or by right-clicking the network icon in the system tray and selecting Test internet speed. While the test runs, Quick Settings shows its progress and results, including download speed, upload speed, and ping.
Because the test runs against your active internet connection, you can quickly understand how your current connection is performing without leaving your workflow. This updated experience builds on the internet speed test we previously introduced for Windows Insiders, bringing the test and its results directly into Windows.
[caption id="attachment_179244" align="aligncenter" width="432"] Internet speed test results in Windows.[/caption]
Thanks,
Stephen and the Windows Insider Program team
There is something iconic about the sticky note on the computer monitor. One reason they work well is their flexibility, durability, and simplicity. Mini Ink attempts to achieve something similar on pen like devices but with an added template level that pairs well with what works for folks who find value in physical planners.
Grids, graphs, day planners, and hourly itineraries are not going away any time soon. Now they are in your tray with Mini Ink. Plan out your week, day, or draw up a quick sketch; all are super easy to do in an app that doesn’t try to be more than it is. Drawing with pen & ink is such a 2d spatial experience and the templates in this app aim to make that experience quicker at hand and easier to be useful.
I’d love to hear what you think of this experiment that uses the latest WinUI bits; please do let me know!
Log in
