AI Weekly: Copilot Connector Write Actions and Claude Fable 5.2

Microsoft confirmed general availability for Copilot connector write actions in Chat and Cowork on November 9. With a compatible connector and the right permissions, you can ask Copilot to create a Jira issue or update a Notion page. The connector’s publisher decides which actions it exposes. Microsoft’s November 9 announcement.

Cowork could already make changes and ask for approval. Microsoft’s Planner and OneNote integrations had rolled out to Frontier in October with write-back and approval. I tried that approval step with Microsoft’s Planner plugin: an email triggered Cowork, which checked for a duplicate, proposed a task with a due date, and stopped for review. The task stayed a proposal. Microsoft’s October roundup.

November 13, 2026 edition. Sources checked through November 13. Last week’s AI Weekly has the Planner and Copilot Studio MCP background.

Microsoft AI updates: this week in 60 seconds

• Copilot connector write actions: Microsoft confirmed general availability for compatible connectors in Chat and Cowork. New write tools ask for approval by default.
• Claude Fable 5.2: rolling out in Cowork and Copilot Studio. GitHub Copilot has a separate rollout and model policy.
• Copilot Studio: Microsoft shared new details about how the harness keeps working after a tool call fails.
• Power Platform: generative pages can call existing Dataverse custom APIs in public preview. November 15 is also a key date for environment creation and Release Planner.
• Copilot Chat scheduling: Microsoft plans to move scheduled prompts to Tasks → Scheduled. The existing scheduling service retires December 31.
• Beyond Microsoft: Google broadens access to Gemini 4 Carbon and releases Hydrogen, Claude adds collaborative artifacts, and ChatGPT opens a beta for business profiles.

1. What general availability means for Copilot connector write actions

Microsoft had already scheduled the connector rollout to start in October. For the capability covered by the November announcement, Copilot uses a federated MCP connector to call a tool in the source app. Atlassian Rovo, for example, can expose an action to create a Jira issue from a meeting transcript. Your Jira permissions still govern where you can create it. Microsoft says support inside Office apps will follow. October rollout notice · Microsoft Learn: supported writes.

New write tools ask first. Expand the confirmation card to inspect the values Copilot plans to send. Allow once makes sense while you’re getting to know a tool. You can also allow that tool for the conversation. Always allow saves approval for that specific federated tool until you reset it or the publisher changes the tool. If you decline, the action doesn’t run. Users can reset approvals under Settings → Sources. Microsoft is also adding an admin center view that lists tools for supported connectors. Microsoft Learn: write actions and approvals.

How to test Copilot connector write actions

I’d start with one ticket in a test project. Check the destination, the assignee and the details before approving. Then open the ticket in the source app. Ask Copilot to process the same email again and see whether it finds the first ticket. A good approval card gives you a chance to catch a mistake. A duplicate check keeps the same request from turning into extra work.

Try declining the request once. Confirm the ticket was never created, then run the same test with approval and check the result in the source app.

Copilot Chat scheduled prompts move to Tasks

Microsoft plans to move scheduled prompts to its native scheduling service. The November 12 notice lists a November 11 start for Frontier and TAP and a worldwide rollout from November 30 through December 31. Eligible users will manage prompts under Tasks → Scheduled in the Microsoft 365 Copilot app on web and Windows. The Power Platform backed scheduler retires December 31. If you rely on a weekly briefing or recurring inbox check, verify it appears in the new view and recreate it if it’s missing. Message Center notice MC1550533.

The Microsoft 365 outage affected connected AI tools

During the November 12–13 Microsoft 365 outage, Microsoft reported problems loading Copilot Chat and using work data to ground responses for affected users. Anthropic reported Microsoft 365 connector failures and linked them to the outage. OpenAI also reported Microsoft 365 connector errors. If a tool fails to retrieve mail or tasks, surface the failed fetch before using the result to make a decision. Microsoft advisory via Penn · Anthropic status · OpenAI status.

2. Claude Fable 5.2 starts rolling out in Cowork and Copilot Studio

On November 11, Microsoft said Fable 5.2 had started rolling out to eligible Cowork and Copilot Studio users. Availability can vary by organization and region, and admins can manage access in the Microsoft 365 admin center. In Copilot, Work IQ gives the model access to relevant work context within the user’s existing permissions.

Anthropic introduced Fable 5.2 and Mythos 5.2 on November 9. They share an underlying model, but differ in safeguards and access. Fable is the broadly available version. Mythos access is gated through Anthropic’s verification programs for eligible cybersecurity and life sciences work. The Microsoft Copilot announcement covers Fable.

The first comparison I’d run in Cowork is a handoff with conflicting dates in two source files. Give both models the same files and ask for an owner, due date and source for each action. I’d score which model catches the disagreement, how much cleanup the result needs, and what the full run costs in Cowork. That tells me more about when to reach for Fable than a better answer to a one-line prompt.

3. What Copilot Studio agents do after a failed tool call

Microsoft’s November 9 Copilot harness deep dive follows the October announcement. The harness is the part that runs the work: take a step, inspect the result, decide what to do next, and repeat within time, step and token limits. It can work with files and run code in a secure container, while Copilot Studio handles connector access and credentials.

Microsoft reports that agents kept working after a failed tool call in 84–91% of observed cases across its autonomous and Microsoft 365 Copilot channels. That measures whether the agent continued. To know if it finished, you still have to check the result. One example from the write-up: Microsoft added read tools to let agents inspect which tables and columns exist in a workbook before trying to use them. That’s a concrete fix for an agent assuming the worksheet already contains an Excel table.

My test would include a workbook with a missing table and a separate record the user can’t access. In the first case, the agent should identify what’s missing and proceed only where it has valid input. In the second, it should report the permission block. An agent that keeps going is useful when it also knows which work it has to leave unfinished.

There’s a parallel change for developers. In Visual Studio 2026 version 18.12, Copilot Chat uses the Microsoft Copilot SDK by default. You can steer a run in progress, inspect step results and see the credits used. In Microsoft’s C# evaluation, the new experience resolved 124 of 157 tasks versus 103 in the old experience, both with GPT-5.6 Sol. That’s a useful reminder to record which version of the editor ran your test alongside the model name.

4. Power Platform: put business logic behind the page

The November 10 Power Platform roundup covers October and November, so some of it will look familiar from PPCC. Here are three items I’d put on a maker’s test list:

• Generative pages and Dataverse custom APIs (public preview): a generated page can call an existing global or table-bound action or function. An Approve order button can call the same Dataverse operation other apps already use. Calls run as the signed-in user, so test the same button with an approver and a user who lacks that permission.
• Generative page telemetry (public preview): if your admin enables Application Insights export for the environment, published pages can report supported errors, load performance and data operations. Try a failed call in a sandbox and see whether the trace gives you enough to investigate.
• Business Applications in Work IQ (preview): test the actual model-driven app form behind the task. Microsoft documents that forms with JavaScript OnLoad, OnChange or OnSave handlers can’t be used for create or update in Work IQ. Those changes still need the form in the app.

Power Platform admin checks

Connector inventory. Microsoft’s November 9 Message Center notice (MC1546312) lists November 10 as the general availability date for connector inventory. Microsoft Learn still labels it preview, so check for the Connectors column under Manage → Inventory in your tenant. I’d use it to find which apps, flows and agents use a connector before changing its data policy.

Copilot Studio model retirements. Microsoft lists November 12 as the retirement date for GPT-4.1, GPT-4.1 Mini and GPT-5 Chat, Auto and Reasoning in affected Standard harness environments. The planned replacement is GPT Chat Latest, with up to 30 more days available by extension. Microsoft lists exceptions for FedRAMP tenants and non-US environments that restrict cross-geo data movement. If an agent switches models, rerun your tests for tool choice, response format and failed lookups. Microsoft Learn retirement table.

Check SharePoint temporary links. A November 13 Power Platform notice flags custom apps and flows that pass SharePoint tempauth links to download files or show thumbnails. Microsoft plans to start retiring these links April 1, 2027. Search custom formulas and flow handoffs for tempauth and test an authenticated replacement where the link is consumed. Message Center notice MC1551122 · Microsoft developer guidance.

5. Two new GitHub Copilot models, with different defaults

GitHub announced general availability for Claude Fable 5.2 on November 9 and Gemini 4 Carbon on November 10. Both are rolling out gradually on Pro+, Max, Business and Enterprise. GitHub says Carbon follows the default model enablement policy. For Business and Enterprise, Fable 5.2 requires an admin to turn its policy on.

GitHub also says Anthropic retains Fable 5.2 prompts and outputs by default to operate its safety classifiers, and offers a zero data retention option for certain eligible enterprises. I’d check that policy before using it with sensitive repositories. For the coding comparison itself, run both models against the same bug and compare the passing tests, changes outside the requested scope, and cost of the whole run.

6. Beyond Microsoft: Gemini 4 Hydrogen, Claude Sheets and ChatGPT business profiles

Gemini 4 Carbon and Hydrogen

Gemini 4 Carbon. Google broadened access on November 10 after a limited cybersecurity release. It’s available through the paid Gemini API and rolling out to Google AI Ultra across supported surfaces. Google positions Carbon for complex, longer tasks and Argon for most day-to-day coding and knowledge work. If you’re budgeting an API pilot, the introductory $2 input and $10 output per million tokens becomes $4 and $20 after November 30. Track how many tokens the full task uses as well as the list price.

Gemini 4 Hydrogen. Google introduced Hydrogen on November 12 for faster coding and agent work. The API price is $1.50 per million input tokens and $7.50 per million output tokens across its one million token context window. Google is also introducing Orchestrator in Antigravity, which can switch models within a session. A harder coding task can use a reasoning model for planning and Hydrogen for implementation and tests. Google announcement · Antigravity availability by plan.

Claude and ChatGPT workspace updates

Claude Sheets, Whiteboards and Data. Anthropic’s November 10 update brings collaborative spreadsheets and whiteboards, dashboard work in Claude Data, and interactive connectors that can open supported partner apps inside Claude. Meetings was announced as coming soon. I’d try one workflow that starts with a shared board and ends with a tracker someone else can edit. The test is whether the next person can pick up the work without having to read the whole conversation.

Businesses in ChatGPT. OpenAI’s November 10 beta lets eligible businesses claim and update a profile and see aggregate data such as mentions and website clicks. Business Agents are a separate part of the launch, initially for eligible US ecommerce and online service businesses. Customers choose to start a labeled conversation with the business’s agent, separate from their original chat. For a Canadian business, checking profile eligibility is a better first step than planning around the agent rollout.

Replit and NVIDIA

Replit Agent can prepare purchases and ads. The November 12 update lets eligible US Core and Pro users prepare one-time purchases through Stripe Link and Visa, with the user approving the final total. Pro users can also prepare Google Ads campaigns and approve them before launch. That puts the amount, the merchant and the approval on the same path as the action. Replit announcement.

NVIDIA named an interim CEO. NVIDIA announced on November 12 that Jensen Huang had resigned. CFO Colette Kress is serving as interim CEO while retaining her CFO role. Harvey C. Jones became board chair. NVIDIA statement.

7. Dates to check before Ignite

• November 15 · Unmanaged Developer environments. New creation ends, including provisioning through the free Developer Plan sign-up. Existing unmanaged Developer environments start being suspended December 15 and deleted January 15. Routed Managed Developer environments are unaffected. In the admin center, filter for Type: Developer and Managed: No, then find the owner and decide what needs to move or be preserved.
• November 15 · Dataverse for Teams and Release Planner. Creation of new Dataverse for Teams environments ends. Existing environments have their own 2027 timeline. The Release Planner transition to the AI at Work roadmap also completes by November 15. Keep using Message Center for changes that affect your tenant.
• November 16 · GitHub Copilot model retirements. GitHub lists Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra. Opus 4.8 has an exception for individual annual subscribers. Check any workflows pinned to these models before Monday.
• November 17–20 · Microsoft Ignite. Ignite runs in San Francisco and online. Microsoft plans to open the public Copilot Plugin Catalog on November 17. It also says Plugins will start appearing in Copilot Home navigation. I’ll be watching for what moves from preview to general availability and how admins can enable it. Microsoft Plugin Catalog announcement.

My first test this week would be a single ticket created through Copilot connector write actions. The approval card tells you what Copilot intends to do. The ticket in the source app tells you whether it did it.