The Week in Agents, 30 May to 5 June 2026

Eight shifts in the Copilot Studio agent platform this week: Frontier Tuning brings reinforcement learning inside your compliance boundary, computer-using agents hit general availability alongside a redesigned workflows canvas, a new Skills capability appears in preview environments, Mistral Medium 3.5 lands as an experimental model with a notably different governance posture, Microsoft Entra Agent ID extends governed identity to Dataverse agent users, Microsoft Scout introduces the Autopilot category, the Dataverse MCP server gives coding agents guardrailed access to your data layer, and Microsoft Research released Webwright as a different way to build browser agents.

TL;DR: Copilot Studio shipped Frontier Tuning (private preview via Forward Deployed Engineers), GA for computer-using agents, a new workflows designer in early release, a Skills capability spotted in preview environments, and Mistral Medium 3.5 as an experimental model. Microsoft Entra Agent ID for Dataverse agent users is in preview, Microsoft Scout (Autopilot) is in Frontier private preview, the Dataverse MCP server is released, and Microsoft Research published Webwright for code-as-action web agents. The governance posture varies sharply: Mistral data is processed outside Microsoft’s managed environments and the standard Microsoft DPA does not apply. Last updated June 2026. Catch up on last week’s edition (24 to 30 May 2026) if you missed it.

Frontier Tuning brings tunable Copilot Studio agents inside your compliance boundary

At Microsoft Build, Microsoft introduced Frontier Tuning, a reinforcement-learning approach to adapting agents to your business. Tuning runs inside a managed Reinforcement Learning Environment used for both post-training and inference. At inference, the system explores multiple frontier and fine-tuned models from Microsoft AI and OpenAI across turns to find stronger candidate paths before returning an answer.

The key promise for governance-aware engineering leaders: tuned models, embeddings, skills, orchestration logic, and the runtime harness all stay within your compliance boundary. Models inherit the access controls of the source data, so only people who could see the underlying content can access models built from it. Tools are virtualized, which means agents can improve without touching production systems.

Availability today: private preview through Microsoft’s Forward Deployed Engineers, with named partners including Land O’Lakes, EY, Bristol Myers Squibb, Pearson, McKinsey, McCarthy Tétrault, and the Josh Bersin Company. General availability inside Copilot Studio and Microsoft Foundry is coming, with no public GA date yet. If you have an FDE relationship, this is the path to start. If not, plan for the Copilot Studio surface to be the entry point.

Read more: Frontier Tuning: Teaching AI to work the way you do on the Microsoft Copilot Studio blog.

Computer-using agents go GA in Copilot Studio, plus a new workflows designer

Computer-using agents (CUAs) in Copilot Studio are now generally available. CUAs interact with websites and desktop applications through the user interface, which closes the gap on systems that lack APIs entirely. The GA release adds enterprise-ready capabilities you actually need to run these in production: more secure credential management, model choice per automation scenario, and adaptive behavior that does not break the moment a vendor portal changes its DOM.

The matching announcement is the new workflows designer, now in early release environments. It gives you a single visual canvas for orchestrating API actions, approvals, business logic, and adaptive UI interactions in the same automation. The interesting primitive is the agent node: a workflow step that hands off to an agent when the decision needs reasoning over context, multiple tools, or knowledge retrieval, then returns control to the deterministic flow.

Practical read: this is the first time the Copilot Studio surface has a credible answer for “I need a workflow that is mostly deterministic but calls an agent at the messy steps.” That pattern was previously stitched together across Power Automate, custom prompts, and topic flows. Now it is one canvas. Plan to revisit any classifier or orchestrator topic you built in the last six months and ask whether it should be a workflow with an agent node instead.

Read more: New and improved: Computer-using agents, a new workflows experience, and real-time voice experiences.

A new Copilot Studio Skills capability is showing up in preview environments

Alongside the redesigned workflows canvas, a new Skills capability is appearing in early-release preview environments of Copilot Studio. Per public demos from Microsoft’s Copilot Studio product team this week, makers can either author a skill directly in the UI or upload a Skill.md file. Tenant administrators can govern Skills from the Power Platform admin center. Demos also show Python execution as part of a skill.

Treat this one as forward-looking. There is no public Microsoft Learn documentation page for Copilot Studio Skills yet, so do not bet a production roadmap on the exact shape of the feature. What it tells you, though, is that Microsoft is moving the agent-extension story toward a portable, file-defined contract that lives next to your code rather than only inside the Studio canvas. If you have ever wished your Copilot Studio agents looked more like a checked-in artifact, this is the direction.

Read more: track the official monthly Copilot Studio update for the formal announcement, and watch the May 2026 release blog for the surrounding authoring changes.

Mistral Medium 3.5 lands in Copilot Studio with a different governance posture

Mistral Medium 3.5 is now available in Copilot Studio for customers in early release environments, marked as experimental. Mistral describes the model as built for long-horizon tasks, reliable multi-tool calls, and structured output that downstream code can consume, with reasoning effort configurable per request. For EU customers, the pitch is in-region data control and one less procurement track to negotiate. If you build EU-facing agents, the same regional sensitivity is why customer-facing flows like MitID need first-class identity and data-residency stories.

Read the governance fine print before you ship anything to production. Per Microsoft Learn, Mistral models are hosted by Mistral outside of Microsoft. Quote: “This data is processed outside all Microsoft managed environments and audit controls, therefore Microsoft’s customer agreements, including the Product Terms and Data Processing Addendum (DPA) don’t apply. In addition, Microsoft’s data residency commitments, audit and compliance requirements, service level agreements, and Customer Copyright Commitment don’t apply to your use of Mistral’s services.” Use of Mistral is governed by Mistral’s own Terms of Service and DPA.

That posture is different from how Microsoft handles its own first-party models or models accessed under the Microsoft DPA. Context: other model providers have been onboarded under Microsoft’s standard data handling over time as the partnerships matured. The expectation should be that this evolves for Mistral too. For now: pilots, evaluations, and non-production scenarios only, exactly as Microsoft recommends. Admins must opt in twice, in the Microsoft 365 admin center and in the Power Platform admin center, before any maker can pick the model.

Read more: Mistral joins Copilot Studio’s growing lineup of model providers, and the Learn page on connecting Mistral AI models.

Microsoft Entra Agent ID extends governed identity to Dataverse agent users

Microsoft Entra Agent ID is the identity framework that gives AI agents enterprise identities you can authenticate, authorize, govern, and audit the same way you do users and workload identities. Agent identity blueprints act as templates for creating individual agent identities with parent-child relationships, so security policies stay consistent across large fleets. The platform supports OAuth 2.0, MCP, and A2A as protocols.

The new piece, currently in preview per Microsoft Learn: agents that need data access to a Dataverse environment can have an Entra agent user created and associated with their identity, after which standard Dataverse security roles assign access to APIs and data. Copilot Studio integrates with Entra Agent ID so agents created there can automatically receive an Entra agent identity that is visible and manageable in the Microsoft Entra admin center.

Why this matters: Conditional Access for agents (Entra ID P1), Identity Protection for agents (P2), Identity Governance for agents (P1), and network controls via Microsoft Entra Internet Access all light up on top of this primitive. If you have agents reading Dataverse via service accounts or shared connection references, you are accumulating identity debt. Plan the migration to agent users now, even if your first move is just an inventory.

Read more: What is Microsoft Entra Agent ID and Agent users in Dataverse (preview).

Microsoft Scout introduces a new Autopilot category for always-on agents

Microsoft introduced Microsoft Scout, the first agent in a new category called Autopilots: always-on agents that work autonomously, hold their own identity, and act on your behalf. Scout is built on OpenClaw open-source technology, integrates across Teams, Outlook, OneDrive, and SharePoint, and is grounded in Work IQ so it builds context on how you actually work over time.

The interesting part for governance leaders is the identity model. Every Scout agent operates under its own governed Entra identity, not a shared service account, so its actions are attributable. Credentials are scoped to the task and redacted from logs and diagnostics. Sensitive actions can require human sign-off, and Microsoft Purview policies (sensitivity labels and data loss prevention) are enforced inline before anything is sent or written. This is the same design pattern as Entra Agent ID for Dataverse, applied to a productivity-tier agent.

Availability is intentionally narrow today: an experimental release through Frontier, requiring Frontier enrollment, Intune policy configuration, and an opt-in attestation. Users with a GitHub Copilot license can then download the desktop experience. If your organization is already in the Frontier program, this is worth piloting on a small group, primarily to learn how always-on, proactive agents change calendar and follow-up patterns, before you decide what to do at scale.

A personal note: Microsoft Scout is the agent I used to rebuild this site. Over a weekend, I had Scout drive the iiu.dk theme refresh end to end, from design exploration to publishing. If you want to see what an Autopilot agent feels like in practice, the play-by-play is in The work as we know it is gone: I rebuilt iiu.dk in a weekend with an agent.

Read more: Introducing Microsoft Scout: your always-on personal agent.

Dataverse MCP server gives coding agents guardrailed access to your data layer

Microsoft released the Dataverse Remote MCP Server at Build, with a public lab repo on GitHub. The repo walks you through three labs: setting up a Dataverse environment with the local MCP proxy, using the server with Claude, and using it with Visual Studio Code and GitHub Copilot. The point is that any MCP-aware coding agent (Claude Code, GitHub Copilot, Codex, custom clients) can now read Dataverse metadata, navigate solutions, and interact with data behind the same security model your tenant already has.

This is a meaningful shift if you do Power Platform development. Coding agents previously had to route through the Web API or custom plugins, which meant either crafting prompts that listed the schema or letting the agent guess. With an MCP server, the agent discovers tables, relationships, and security context as a tool surface. The guardrails come from the Dataverse environment itself, not from prompt engineering.

Read more: microsoft/Dataverse-MCP on GitHub and the Dataverse MCP Server User Guide.

Microsoft Research’s Webwright reframes browser agents as code, not click streams

Microsoft Research, with collaborators at the University of Hong Kong, published Webwright: a deliberately minimal terminal-based harness that lets a coding agent solve web tasks by writing Playwright code rather than predicting one click at a time. Three modules, roughly 1,000 lines of code, one agent loop, no multi-agent orchestration. Per the research blog, this setup reaches state-of-the-art results on Odysseys and Online-Mind2Web with a 100-step budget.

The reusable artifact is the program, not the browser session. Once the agent has solved a task, the resulting script can be packaged as a CLI with arguments and reused across platforms (the post mentions Codex, Claude Code, and OpenClaw as runtimes). Cost analysis: GPT-5.4 averages $2.37 per task, with a smaller Qwen3.5-9B model still hitting strong performance on the hard split of Online-Mind2Web given the right tools.

This is research, not a product, but it is research worth tracking. If you maintain RPA flows or Playwright-based automations, the Webwright pattern (agent writes code, browser is disposable, script is the artifact) is a credible alternative to click-recording. Watch for it to influence the next round of computer-using agent harnesses.

Read more: Webwright on the Microsoft Research blog.

What to do this week if you’re building

1. Audit any non-deterministic topic flow. If you have Copilot Studio topics that orchestrate multiple tools or branch on reasoning, prototype the same logic in the new workflows designer with an agent node at the messy step. Compare maintainability, not just behavior.
2. Inventory anything UI-driven and brittle. Computer-using agents are now GA. List the legacy portals and desktop apps your team automates with scripts, RPA, or “please copy-paste this.” Pick one for a CUA pilot, with credential management and a real eval set from day one.
3. Decide your Mistral position before makers ask. Read the Microsoft Learn governance page, decide whether your tenant should opt in, and write down the production-versus-pilot rule. Do not let Mistral show up in the model picker by accident.
4. Map your agent identities to Entra Agent ID. List every Copilot Studio agent and Power Platform automation that touches Dataverse. Identify which ones currently use shared service principals and plan the move to agent users so your security and audit story stays clean.
5. Watch Frontier Tuning. If you have a workflow where eval scores plateau because the model lacks domain depth, flag it as a Frontier Tuning candidate. Talk to your Microsoft account team about FDE engagement, or wait for the Copilot Studio surface and prepare your data.
6. Pilot Microsoft Scout if you are in Frontier. If your organization is enrolled in Frontier and runs Intune, an opt-in pilot on a small group is the cheapest way to learn what always-on Autopilot agents do to calendar and follow-up dynamics. Do not wait for general availability to form an opinion.
7. Stand up the Dataverse MCP server in dev. If your Power Platform team uses GitHub Copilot, Claude Code, or any MCP-aware coding agent, run the GitHub lab. The cost is an afternoon. The payoff is a coding agent that no longer hallucinates your schema.
8. Read the Webwright post. Even if you do not change your stack, the code-as-action framing is a useful lens for evaluating any browser-agent product pitch you receive in the next quarter.