1. Copilot Studio replaced its orchestrator
On 3 August, Workflow Designer went GA — one canvas where you can mix deterministic steps with agent reasoning nodes, test node by node, and get proper version control. That last part matters more than it sounds. Most of the failed pilots I've seen died from an unreviewable change, not a bad model.
The bigger change is underneath: Microsoft has swapped the old topic-based conversation model for an agent-first orchestrator. Their own numbers say roughly 20% better task completion and about half the token spend. I'd treat both figures as a starting point for your own measurement, not a number to put in a business case unverified — but directionally, an agent-first router doing less re-explaining per turn is the right shape of improvement.
Computer-using agents — GA since May — can now sit inside a workflow as a node, so browser and desktop automation runs inside the same orchestrated process as everything else. That's the RPA-to-agent bridge a lot of enterprise clients have been waiting for before they'll commit spend.
Run-only sharing entered preview in August too: you can hand an agent to end users to trigger and monitor without exposing the logic that built it. GA is targeted for January 2027. For anyone doing agent governance, this is the feature that finally separates "who can run this" from "who can edit this" — a distinction Copilot Studio has needed since day one.
The distinction between who can run an agent and who can edit one is the governance question every serious rollout eventually asks.
2. Agent Builder tidies up, quietly
Late August brings three small changes to Microsoft 365 Copilot's Agent Builder: new agents get default capabilities similar to core Copilot out of the box (still fully editable), the knowledge-source setup gets a redesign, and "Uploaded Files" becomes "Attachments." None of it is dramatic. All of it lowers the floor for a business user building their first agent, which is exactly where Agent Builder should be aiming — it's the on-ramp, not the destination. Copilot Studio is where you graduate to when the agent needs real orchestration.
3. Cowork is generally available, and now it has a bill
Copilot Cowork — the agent that runs a multi-step task to completion rather than handing you a draft — went GA on 16 June after a three-month preview that reportedly involved more than half the Fortune 500. It's licensed through Microsoft 365 Copilot, but usage is metered separately through Copilot Credits, priced against model use, context retrieval, tool calls, and runtime.
Two details matter for anyone running the numbers. First, it's disabled by default — an admin has to turn it on, and can cap spend at tenant, group, or user level. Second, the billing model means the business case for Cowork isn't a flat licence cost, it's a consumption forecast. If you're building an ROI case for Cowork, model your credit spend against a real task sample before you present a number to finance, not after.
4. Scout is the always-on one, and it comes with an identity
Scout, announced at Build in June, is a different animal from Copilot and Cowork: it doesn't wait to be asked. It watches Teams, email, calendar and files on an ongoing basis, drafts responses, resolves meeting clashes, and chases commitments that were made in passing and never written down anywhere formal.
What I'd flag to anyone evaluating it is the governance layer, because it's unusually well specified for a first release. Scout runs under its own Entra agent identity, needs Intune policy configuration, and every action it takes is attributable and auditable. That's not incidental — an agent that acts without being asked is the one your compliance team will ask about first, and Microsoft has clearly built the answer in ahead of the feature rather than after. It's currently gated behind the Frontier program and a GitHub Copilot licence, so treat it as early rather than ready for a general rollout.
5. Foundry keeps building the plumbing underneath all of it
The June Foundry update is where the infrastructure story lives. Claude reached general availability on Foundry, hosted on Azure with the full Messages API — so "which model" is now genuinely a build-time decision rather than a platform lock-in. Agent publishing straight to Microsoft 365 Copilot and Teams also went GA, meaning one governed pipeline instead of a separate rebuild per surface, which is the kind of unglamorous change that saves the most engineering time.
Autopilot agents entered public preview — agents with their own Entra Agent ID, licence, email, calendar and Teams presence, the reference example being a Workstream Manager for group chats. Toolboxes picked up Skills, Work IQ, Fabric IQ, browser automation and tool search, the last of which exists specifically so a large toolbox doesn't have to describe every tool on every turn. And memory got production features — procedural memory, time-to-live controls — which is the unglamorous plumbing that decides whether an agent's context still makes sense after six months in production.
What I'd tell a client this week
None of this changes the fundamentals. An agent still needs a narrow, well-defined job, a governance story before launch not after, and a business case built on a real task sample rather than a vendor number. What's changed is that Microsoft has finally built separate lanes for the different kinds of agent work — Agent Builder for the quick win, Copilot Studio for the orchestrated build, Cowork for the metered multi-step task, Scout for the always-on watcher — and given each one its own identity and audit trail. That's the part worth paying attention to. The rest is detail.