Microsoft Copilot Studio changed substantially in 2026, moving well beyond conversational chatbots toward a broader platform for long-running, action-oriented business agents. If you evaluated it twelve months ago and moved on, the product you would encounter today is meaningfully different — in what agents can do, how they are governed, and what they cost.
This briefing covers the most consequential changes from January through mid-November 2026 and translates them into decisions that matter for business leaders rather than platform engineers.
The one-sentence summary: Copilot Studio evolved from a conversational agent builder into a platform where agents can operate software, collaborate with other agents, hold persistent identities, and be built alongside traditional apps — with governance and billing complexity to match.
The defining change: a new runtime architecture
The most consequential shift in 2026 is the new Microsoft Copilot harness — a rebuilt reasoning and runtime environment designed for longer, multi-step work. Microsoft first released it as a production-ready preview on June 9, promoted it to general availability on August 3, and renamed it from the "GitHub Copilot harness" to the "Microsoft Copilot harness" in late October.
What this means in practice: agents built on the new harness can reason through complex, multi-step tasks rather than relying on scripted conversation branches. Builders can describe what they want in natural language and the harness handles more of the orchestration. The standard harness still exists and remains appropriate for more deterministic scenarios.
The important caveat for business decision-makers: the new harness carries usage-based billing. Unlike standard-harness agents, Microsoft Copilot-harness activity can consume Copilot Credits while the agent is being built and tested — not only after it is published to users. Organizations need to treat harness selection as an architecture and cost-management decision, not a maker-level choice.
What agents can now do
Three capability expansions materially change what a Copilot Studio agent can accomplish:
Computer use (GA since May 13 — standard harness). Agents can now operate web browsers and Windows desktop applications through their UI — filling forms, navigating screens, extracting data — without a purpose-built API. This is particularly valuable for systems that have no API and are unlikely to get one. One important qualifier: computer use is currently supported on the standard harness; Microsoft's feature matrix lists it as planned, not yet available, on the new Microsoft Copilot harness. The practical constraints also include controlled credentials, appropriate machine infrastructure, and metered Copilot Credit consumption — billed per model step (5 credits for a standard-model step, 15 for premium), not per individual click or keystroke.
Multi-agent collaboration (A2A GA since April). Copilot Studio agents can delegate work to other specialized agents using the open Agent2Agent (A2A) protocol. By October, preview capabilities also allow a Copilot Studio agent to be exposed as a service to external clients and orchestrators via MCP and A2A channels. The practical implication is that you can build a network of specialized agents rather than a single monolithic one — but this broadens the trust boundary that administrators need to govern.
Work IQ and Microsoft 365 grounding (GA October; default for new Microsoft Copilot-harness agents). Agents built on the new Microsoft Copilot harness can now answer and act using permitted context from a user's email, calendar, files, Teams messages, and people data. This eliminates a significant amount of custom integration work for agents that need to operate in the context of an employee's actual work. Two defaults matter for governance: all available workloads are enabled by default, and write operations are on by default (subject to organizational permissions and policies). Workload selection can narrow which services an agent sees, but cannot target individual mailboxes or sites within a workload. Some workloads remain preview features. September also added Dataverse as a native knowledge source, and October added OneDrive for Business knowledge with Purview sensitivity-label support.
Persistent agents: the preview to watch
The most forward-looking development, currently in Frontier preview (October), is persistent agents with their own accounts. Under this model, an agent can have its own Microsoft Teams presence, an Exchange mailbox, memory across sessions, and event-driven triggers. Rather than waiting for a user to invoke them, these agents can initiate action when something happens — a new email, a calendar event, a data change.
This matters strategically because it moves agents from tools you use to colleagues you delegate to. The governance implications are significant: persistent agents with their own identities and event subscriptions are first-class principals in your environment, not just software features. Organizations that begin thinking about agent lifecycle management, permissions, and accountability now will be better positioned than those treating this as an IT detail.
Before exploring this preview, note the eligibility requirements: access requires an early-release environment, tenant-level Frontier enrollment, administrator activation, eligible maker licensing, and a per-agent Microsoft 365 Frontier for AI Teammates license. Account licensing and consumption charges are separate items. Current functional limitations include no support for interactions with external users, and restrictions around protected documents. Ordinary Copilot Studio access does not establish access to agent accounts.
Key milestone table
| Capability | Status | Date |
|---|---|---|
| Microsoft Copilot harness (new runtime) | GA | August 3, 2026 |
| Computer use (UI automation) | GA | May 13, 2026 |
| Agent-to-agent (A2A) connections | GA | April 2026 |
| Agent evaluations (core) | GA | March 2026 |
| Evaluation automation (REST/API) & custom analytics metrics | Preview | Apr–Sep 2026 |
| Microsoft Entra Agent ID (mandatory) | GA | Rollout from May–July 2026 |
| Work IQ (M365 grounding, default for new Microsoft Copilot-harness agents) | GA | October 2026 |
| Apps inside Copilot Studio | Preview | September 2026 |
| External MCP/A2A publication channels | Preview | October 2026 |
| Persistent agents / agent accounts | Frontier | October 2026 |
Copilot Studio is becoming a business-solution environment
September's announcement that apps are now available inside Copilot Studio (public preview) signals something important about Microsoft's product direction. Makers can now describe an application in natural language, receive a generated implementation, and combine that interface with workflows and agents — all within one studio. These apps run on Microsoft Copilot Managed Runtime, Microsoft's managed hosting layer.
The practical implication: Copilot Studio is no longer purely an agent editor. It is becoming a business-solution construction environment that can produce agents, workflows, and app interfaces together. The licensing picture has two distinct parts: building an app in Copilot Studio is charged through Copilot Studio billing, while running those apps can be covered either by Power Apps Premium licenses (which exempt users from runtime Copilot Credit consumption) or by a spending policy that bills credits. The app experience and Managed Runtime remain preview technologies, so production commitments should wait for GA.
Governance and administration: now a first-class concern
Microsoft Entra Agent ID is now mandatory for new agents (rollout from May–July 2026, with some documentation discrepancy on the exact date). Each agent automatically receives a managed directory identity — a directory object that counts against Entra tenant resource quotas. For organizations with large agent estates, this creates a tangible identity-lifecycle consideration: agents need to be provisioned, governed, and deprovisioned like other principals.
October also gave Power Platform administrators a direct control over which knowledge-source types makers may use — covering files, SharePoint/OneDrive, and public websites. This provides a simpler policy boundary around grounding sources, complementing existing data-loss-prevention controls.
For quality governance, core agent evaluations moved to GA in March. The September Review panel (also GA) surfaces agent-readiness issues before publication. REST-based evaluation automation and custom analytics metrics arrived later but remain in preview — they should not be treated as generally available release-gate tooling yet. Together, these features make formal pre-production evaluation significantly more practical than it was at the start of 2026, though the preview components require that caveat.
One important contraction
Amid the expansion, one capability was cut: new WhatsApp connections through the Azure Communication Services-based channel stopped being available on October 23, 2026. Existing connections created before that date continue to operate. Organizations that planned to build new WhatsApp-channel agents through Copilot Studio should evaluate alternative approaches.
What this means for your organization
The 2026 changes push Copilot Studio into territory where the technology decisions overlap significantly with organizational decisions. Here is where business leaders — not just IT — need to be in the conversation:
- Harness selection is an architecture decision. Choosing the Microsoft Copilot harness for new agents has billing implications that differ from the standard harness. Credits can be consumed during development, not only production. Microsoft Copilot harness usage is billed on Copilot Credits regardless of whether users hold a Microsoft 365 Copilot license — the two are independent. Cost controls and spending policies should be in place before pilots begin.
- Agent identity governance is now real infrastructure. Agents with Entra identities, persistent accounts, and event-driven triggers are principals in your tenant, not just software. Establish lifecycle management for agent identities before the estate grows.
- Computer use opens legacy systems — with caveats. UI automation for systems without APIs is genuinely useful, but it requires careful thinking about credentials, auditability, and what happens when the UI changes. This is a capability worth piloting in controlled conditions, not deploying broadly without an operating model.
- The preview roadmap signals where Microsoft is investing. Persistent agents, app generation, and external MCP/A2A channels are all in preview. Organizations that want to influence Microsoft's direction — or simply be ready when these capabilities become GA — should be testing them now under appropriate governance.
- Grounding is now much easier, but the permission model differs by agent type. For standard Work IQ agents, the agent sees content the invoking user is permitted to see — so your M365 permission hygiene becomes the agent's permission hygiene. Frontier persistent agents operate differently: they use their own agent account identity and scoped access rights rather than inheriting a single user's permissions. Critically, the shared memory these agents maintain has no privacy boundary between users — summaries of files or conversations can surface in responses to people who could not access the original content. Users cannot opt out of shared memory in the current preview. Treat this as a material data-governance consideration before testing persistent agents with sensitive content.
The bottom line: Copilot Studio in late 2026 is a significantly more capable and significantly more complex platform than it was twelve months ago. The capabilities that matter most for business outcomes — computer use, multi-agent collaboration, persistent agents, and richer grounding — are now real or approaching general availability. The governance, cost, and identity infrastructure to support them has not always kept pace. Organizations that invest in that foundation now will capture the upside; those that treat it as a detail will find it catching up with them.
Making sense of AI platforms for your business
Not sure which capabilities are ready for your use cases, or how to build a governance foundation for enterprise agents? We work with business leaders on exactly these decisions.
Talk to us