The Week in Agents, 8 to 14 June 2026

You are still opening the old Copilot Studio. The rebuilt version, with a new agentic orchestrator, nine tabs reduced to four, and Skills support, went generally available on June 9. Here is what changed in Copilot Studio and Dataverse since the last newsletter, plus what you should action this week.

TL;DR: Copilot Studio: new experience went GA June 9 (new orchestrator, Skills, 4-tab editor), xAI Grok 4.1 Fast joined the model lineup with the same outside-Microsoft-DPA posture as Mistral, agent memory is showing up in demos (watching), and Ricardo Calejo’s SkillWright demo shows where the Skills ecosystem is heading. Dataverse: low-latency sync to Fabric is reportedly GA as of June 10 (official blog not yet located), and the Dataverse MCP server picked up Skills as tools. Catch up on last week’s edition (1 to 7 June 2026) if you missed it. Last updated June 2026.

Copilot Studio this week

New Copilot Studio is generally available. Last week I flagged the new Copilot Studio as a preview I was watching. On June 9, it went generally available worldwide. The announcement surfaced in my feed via Velmurugan Muthaiyan and Henry Jammes.

The core change is the orchestrator. The new one is built on a coding harness and CLI layer, which Microsoft says produces stronger instruction adherence and long-horizon task execution. It supports recursive task execution, meaning agents can work through complex and dynamic problems step by step rather than failing partway. Nine configuration tabs are now four: Instructions, Skills, Tools, and Knowledge on one surface. Inline chain-of-thought and tool-call visibility are part of the test experience. Skills are GA: write reusable instructions in markdown that load on demand, or import existing GitHub Copilot and Claude Code skills. The workflow designer ships alongside: visual canvas for agentic automations, agent nodes, node-by-node testing, and MCP server connections (last one still preview). The old experience stays. Entry point: “Try now” at the top of the Copilot Studio homepage.

Read more: Meet the new Copilot Studio: rebuilt for more complex, multi-step work, June 9, 2026.

xAI Grok 4.1 Fast joins the model lineup, same DPA pattern as Mistral. xAI Grok 4.1 Fast is now available in preview in early access environments. US-based makers, admin opt-in required (M365 admin center, then Power Platform admin center). Fast reasoning, large context, deep tool use.

From the Microsoft blog: “xAI’s models are hosted outside Microsoft-managed environments, and when you use Grok 4.1 Fast in Copilot Studio, your relationship with xAI will be independent of Microsoft and governed by xAI’s Enterprise Terms of Service and Data Processing Addendum.” Same posture as Mistral from two newsletters ago. Customer data is not retained or used to train xAI’s models, per Microsoft, but the standard Microsoft DPA, data residency commitments, and Customer Copyright Commitment do not apply. This is the second external-DPA model provider in a month. Decide your policy once, enforce it uniformly. If Mistral was non-production only, Grok 4.1 Fast is too.

Read more: More choice, more flexibility: xAI Grok 4.1 Fast now available in Microsoft Copilot Studio.

Agent memory in Copilot Studio (watching). Matthew Devaney posted this week that memory has come to Copilot Studio: agents can remember interactions, workflows, and context across sessions. This would be a significant capability shift, turning stateless agents into something that accumulates context over time. No Microsoft Learn page or official blog post confirming the feature yet. If you have early access and see the memory option, I would love to know what the configuration surface looks like.

Dataverse this week

Dataverse to Fabric low-latency sync reaches general availability (reported, primary blog pending). The Microsoft Dynamics 365 Community channel (an official Microsoft-run LinkedIn presence) announced on June 11 that low-latency sync for Dataverse to Microsoft Fabric is now generally available for customer engagement apps, finance and operations apps, and custom Dataverse solutions. This replaces batch-interval replication with near-real-time sync. I could not locate the corresponding Microsoft blog post or updated Learn page. The signal is from an official Microsoft channel so confidence is high, but I am treating it as “reported GA, primary source pending.” If you have a Dataverse-to-Fabric link configured, check whether low-latency sync is now available in your tenant, and validate your analytics pipelines against the faster cadence before enabling it on production data.

Dataverse Business Skills combined with the MCP server is the Dataverse story I keep coming back to this week. Isha Kapoor flagged an upgrade to the Dataverse MCP server that adds Business Skills as tools in the tool surface. The Dataverse MCP server was covered in detail in the 30 May to 5 June newsletter. But Business Skills themselves have had their own Microsoft blog post since May 1 that I had not given enough attention, and the combination is worth explaining properly.

Every organization has processes that live in people’s heads: how to qualify a lead, how to approve a discount, how to onboard a vendor. Microsoft calls capturing these as Business Skills in Dataverse. You write your process once in plain natural language, store it centrally in Dataverse, and agents discover and follow it at runtime. When multiple agents reference the same skill, they follow the same process. Update the skill once, and the change applies everywhere, with no need to patch individual agent configurations. Skills are solution-aware, which means they travel with your ALM process across environments like any other Dataverse artifact.

What the MCP server upgrade adds is that any MCP-aware coding agent can now invoke these Business Skills directly. Before, a coding agent with Dataverse MCP access could inspect your schema, query tables, and act on data. Now it can also execute your organization’s documented domain knowledge as part of the same operation. GitHub Copilot, Claude Code, a custom Azure AI Foundry agent, and a Copilot Studio agent can all follow the same business process, defined once in Dataverse, without any of them needing their own copy of the logic baked into their instructions.

The practical shift for Power Platform developers: business rules no longer have to live in multiple places (prompt instructions in Copilot Studio, custom actions in Power Automate, a Wiki, a senior developer’s head). One definition, accessible from every tool in your stack. There is a sample business skills repository on GitHub at aka.ms/DVBusinessSkillRepo with production-ready examples you can install directly.

Read more and see it in action: Business skills: Process knowledge that agents can follow on the Power Platform blog. Also: Business skills overview on Microsoft Learn.

What every Power Platform developer and admin should action this week

1. Developers: try the new Copilot Studio today. Click “Try now” on the Copilot Studio homepage. Run one existing agent through the new interface. Test whether the new orchestrator handles your multi-step flows better. Old experience stays if you need to roll back.
2. Developers: put at least one skill in source control this week. Skills are GA. Write one SKILL.md for a repeated task, check it into your repo, confirm the import path. Start treating agent behavior as code.
3. Admins: update your external model policy to cover xAI. Same rule as Mistral: non-production only, admin opt-in. Decide your posture for outside-Microsoft-DPA models once, apply it to every provider that follows this pattern.
4. Admins: check Dataverse to Fabric sync status in your tenant. If you have a Dataverse-to-Fabric link, check whether low-latency sync is now available. Validate analytics pipelines against the faster cadence before enabling on production data.