Microsoft just shipped a rebuilt Copilot Studio, and it's generally available worldwide for production agents. Not a preview. Not a feature flag sitting behind an admin approval you've been chasing for three weeks. If you build agents for a living, this is the kind of release that quietly changes how you work for the next two years. So let's skip the launch-day adjectives and get into what actually moved.
The headline isn't the UI, it's the orchestrator
Sure, the authoring experience got a facelift. Configuration tabs dropped from nine to four, and instructions, knowledge, tools, skills, and model now live on a single Build surface instead of being scattered across the product. That's a genuine quality-of-life win. Onboarding a new consultant now takes an afternoon instead of a week, which matters more than it sounds like it should when you're trying to scale a practice.
But the real change is underneath, in a redesigned orchestration runtime that replaces the classic model outright. It follows instructions more reliably, handles recursive task execution, chews through larger volumes of content, and produces richer file outputs. Microsoft is reporting roughly a 20% lift in evaluation performance alongside close to a 50% drop in net token consumption. If that holds up in real tenants, agents get better and cheaper in the same release, which almost never happens.
The catch: orchestration isn't configurable anymore. Everyone gets the enhanced model, full stop. For most teams that's a relief. If you were one of the few hand-tuning classic orchestration behavior, that lever is gone.
Skills are the sleeper feature
Makers can now write reusable instructions as Skills in plain markdown, loaded on demand for specific tasks. The detail worth sitting up for: you can import skills you already have, including existing GitHub Copilot and Claude Code skills.
That's a real portability story. The work you've already put into shaping agent behavior in one ecosystem doesn't need to be rebuilt from scratch to live inside Copilot Studio. For anyone running agents across Copilot Studio, Claude Code, and GitHub Copilot, which is increasingly everyone, this lowers the cost of standardizing on one skill library instead of maintaining three different dialects of the same instructions.
Workflows finally grow up
The new workflow designer gives you a single visual workspace for building agentic automations, with node-by-node testing and real versioning. The part that matters architecturally is agent nodes: you can call an existing agent directly into a workflow. That means you can mix deterministic, structured steps with agentic reasoning in one process, the predictable parts stay predictable, and the agent handles the judgment calls. It's the pattern most of us have been faking for a year by stitching Power Automate to Copilot Studio with duct tape. Now it's first-class.
Add multi-agent collaboration and open agent-to-agent protocols, and the platform is clearly aiming at orchestrating end-to-end business processes rather than bounded, single-shot tasks. Computer-using agents are in the mix too, driving websites and desktop apps through the UI for the systems that never shipped an API, which, if you've spent any time around legacy ERP environments, is more systems than anyone would like to admit.
Governance kept pace, and that's the surprising part
Capability releases usually outrun their controls. This one didn't. You get clearer agent lifecycle visibility (whether an agent is generating, ready to test, published, or stuck), preview Entra agent identities for access control, and set-level grading with custom metrics on the Evaluate tab. The lifecycle mirrors how agent development actually works: Create, Build, Test, Publish, Monitor. And the Monitor tab, which shows recent tasks and exactly which files an agent touched, is the kind of detail that makes a security review go smoothly instead of sideways.
The edge case to flag before you start anything
Here's the line I'd put in bold in any kickoff deck: agents do not move between the classic and new experiences, in either direction. You pick the experience at creation time, and that's a one-way door. An agent built in classic can't be lifted into the new experience, and vice versa.
So before you start anything new, decide on purpose. For greenfield work, default to the new experience, that's where the orchestrator, the skills, and Microsoft's investment are all pointed. For existing classic agents carrying real production load, plan a rebuild rather than expecting a migration button, because there isn't one. Budget for it now, while it's a line item you control, not later when it's a surprise on someone else's project timeline.
Where we'd start
If you're running a Copilot Studio practice, or you've got agents embedded in a broader D365 or Power Platform footprint, three moves are worth making now:
Rebuild one representative agent in the new experience and measure the token and quality deltas against your classic baseline. Don't take the 20%/50% figures on faith. Verify them on your own data, your own prompts, your own usage patterns.
Consolidate scattered prompt-and-instruction sprawl into markdown Skills you can reuse across Copilot Studio and Claude Code, instead of maintaining separate versions of the same logic in three places.
Inventory your classic agents and tag each one "rebuild," "retire," or "leave it." The clock on classic is running, even if Microsoft hasn't told anyone exactly how fast.
The platform finally matches the way people have actually been building. That's worth more than a redesigned tab, but it also means the gap between teams who plan for it and teams who get surprised by it is about to get wider.
This is exactly the kind of shift where a second set of eyes pays for itself. If you're weighing whether to rebuild a classic agent or bring net-new AI capability into your D365 environment, that's squarely what our work is for, not chasing every capability Microsoft ships, but figuring out which ones actually solve a problem you have.
Reach delivers Microsoft D365 implementations and migrations, and we step in when others can't. Battle-tested across eight industries, whether we're guiding a fresh implementation or rescuing one that's gone off course, we bring deep expertise, transparent communication, and genuine collaboration to every engagement.
