The Week in Agents, 24 to 30 May 2026

Four things moved in the Microsoft agent platform between 24 and 30 May 2026: Mistral Medium 3.5 landed in Copilot Studio as an admin-gated experimental model, Dataverse agent users backed by Microsoft Entra agent identities entered public preview in May with general availability planned for June, Copilot Studio shipped agent nodes inside workflows and workflows as tools inside agents, and Microsoft Research released Webwright, a code-as-action browser agent framework.

TL;DR: Multi-model Copilot Studio is real, but Mistral still runs under Mistral’s own DPA, not Microsoft’s, until it onboards as a subprocessor the way Anthropic did. Agent identity finally gets a first-class home in Dataverse. The hybrid AI plus deterministic workflow pattern has tooling. Browser agents move from action-by-action to script-as-artifact. Last updated May 2026.

Copilot Studio is now genuinely multi-model

Mistral Medium 3.5 is now selectable in Microsoft Copilot Studio for customers in early release environments, labeled as Experimental. It joins OpenAI GPT and Anthropic Claude in a model lineup that has grown noticeably over the last few months.

Three things make this more than a checkbox release.

Admins gate access with a deliberate two-step opt-in: enable Mistral at the tenant level in the Microsoft 365 admin center, then turn on external model providers per environment in the Power Platform admin center. Until both switches are on, the model is not visible to makers. That gives IT a structured path to pilot, evaluate, and expand usage on their terms, the same posture you already use for connector governance and that came up sharply in the recent AI Builder credits change.

Mistral is a European model, which makes the “right model for this scenario” conversation noticeably easier for EU financial services, healthcare, and public sector teams thinking about model provenance alongside data residency.

One nuance worth flagging, and one I expect to change

For now, when an admin opts into Mistral, the model runs on Mistral’s infrastructure outside Microsoft-managed environments. In Microsoft’s own classification, Mistral is an AI independent processor, not a Microsoft subprocessor. That means, as of May 2026, Microsoft’s Product Terms, Data Processing Addendum, data residency commitments, SLAs, and Customer Copyright Commitment don’t apply to Mistral calls. Use is governed by Mistral’s own Terms and DPA, which admins accept explicitly in the Microsoft 365 admin center before the model lights up for makers.

The reason I’d call this a “for now” state: Anthropic was in exactly the same position not long ago. Microsoft initially offered Claude under Anthropic’s separate commercial terms, then onboarded Anthropic as a Microsoft subprocessor. At that point Microsoft Product Terms, the DPA, Enterprise Data Protection, and the Customer Copyright Commitment all extended to Claude usage inside Copilot Studio and M365 Copilot. The Mistral integration looks like the early phase of the same trajectory. I would expect a similar move to the Microsoft-subprocessor framework over the next few quarters, especially given how central data sovereignty has become to EU enterprise buying decisions.

Net for builders: pilot Mistral in non-production scenarios now, document the data-flow assumptions, and be ready to revisit the contractual posture when Microsoft tightens the wrapper.

Read more: Microsoft Copilot Studio blog, Mistral joins Copilot Studio’s lineup · Microsoft Learn, Connect to Mistral AI models · Microsoft Learn, Anthropic as a subprocessor (the precedent to watch)

Agents finally get real identities: Dataverse plus Microsoft Entra agent identity

Dataverse provisions AI agents as first-class agent users,each backed by a Microsoft Entra agent identity, with least-privileged security roles based on the agent’s intended capabilities. Public preview started in May 2026, with general availability planned for June 2026.

Why this matters more than it sounds:

• Agents stop borrowing human credentials. Today, the easiest way to give an agent access to data is to run it as a service principal or, worse, a real user. Both options are bad: one over-grants, the other muddies audit logs. Dedicated agent identities fix both at once.
• Least-privileged, capability-based roles. The agent only gets what its declared task requires. This is the missing primitive that makes agent fleets governable, and makes “this agent went rogue” a debuggable event instead of an existential one.
• Full audit trails per agent. Every Dataverse operation an agent performs is attributable to that agent, not the developer who deployed it. For regulated workloads (financial services, healthcare, public sector) this is the difference between “compliant” and “we’ll get back to you”.

Identity has been the unsolved seam in enterprise agent architectures for the last 18 months. This closes it.

Read more: Microsoft Learn, Create Dataverse agent users with Microsoft Entra agent identity (2026 Wave 1 release plan)

Agents and workflows, finally composable in Copilot Studio

Copilot Studio shipped two new composition patterns this week.Workflows can now call agents through a dedicated agent node (“Add an agent” step), and agents can call workflows as tools, either pre-existing flows from the library or new ones built with natural language inside the agent.

The architectural point: agents bring reasoning and adaptability, workflows bring structure and consistency. Until now you picked one. Real business processes need both, in the same flow, with explicit hand-offs between AI judgement and rule-based execution.

The launch examples make the pattern concrete. A procurement workflow can route to an agent that evaluates vendor proposals against company policy, then return control to the deterministic flow for PO creation and approval routing. An HR onboarding workflow can call an agent to personalize welcome materials by role and department. On the other side, a sales agent can assemble deal context, then call a workflow to generate the quote, apply discount rules, and route it for approval.

The pattern emerging across enterprise agents in 2026 is hybrid: deterministic where compliance demands it, model-driven where flexibility creates value, with explicit hand-offs between them. This is the tooling catching up to that pattern.

Read more: Microsoft Copilot Studio blog, Automate business processes with agents plus workflows

Webwright dropped, and browser automation just got less fragile

Microsoft Research, with the University of Hong Kong, released Webwright, a browser agent framework that takes a fundamentally different approach from Playwright-based agents. The framing in their post is sharp: “A terminal is all you need for web agents.”

The standard pattern today is action-by-action: agent reads the DOM, emits a click, waits, reads again, emits a type, waits, screenshots, retries. Every step is a fragile coupling between the LLM and the page. One layout drift and the run dies in a way that’s hard to debug.

Webwright moves the unit of work up a level. The agent writes a full Python script for the task, runs it in a disposable browser, and the persistent artifact is the code and logs in the local workspace, not the browser session.

Why this is the right shift:

• Code-as-action is debuggable. A Python script is reviewable, version-controllable, and editable by a human. A step log is not.
• Disposable browsers eliminate session drift. Every run starts clean.
• Visual verification matches how humans check work. Real users don’t read the DOM, they look at the screen.

I have been using it in my own setup, including to scrape the LinkedIn activity behind this newsletter, and the difference between “fragile demo” and “saved Python script with screenshot evidence” is exactly as material as it sounds. Related debugging story: fixing autonomous agents stuck on ContentFiltered in Copilot Studio.

Read more: Microsoft Research, Webwright: A Terminal Is All You Need For Web Agents · github.com/microsoft/Webwright

What to do this week if you’re building

1. Read Mistral’s DPA, then run the two-step opt-in (Microsoft 365 admin center, then Power Platform admin center on selected environments) so you can pilot Mistral Medium 3.5 in non-production scenarios.
2. Plan your agent identity migration. If you have agents running on service principals or shared accounts, map them to Entra agent identities now while public preview is live, ahead of GA in June.
3. Pick one workflow where you have been forced to choose between deterministic and AI-driven. Add an agent node where the judgement lives, and keep the structured steps deterministic.
4. Evaluate Webwright against any Playwright-based agent you have in flight. If it would have caught your last three failures, migrate.