The roundup covers more than that, though. It also includes Copilot Studio skills in Cowork, a new Model Context Protocol (MCP) server for developer tools, connector setup inside the chat, a customer conversation harness, Agent2Agent (A2A) interoperability, general availability for Work IQ, three workflow features, and new controls for Copilot Credit spending. Below we cover each one, note its release status, and explain what admins should check before anything reaches production.
The headline: agents that are persistent
Microsoft's blog describes three building blocks:
- Agent accounts. People can talk to an agent and hand it work by @-mentioning it in Microsoft Teams, Outlook, and Office files, the same way they'd tag a coworker.
- Approved credentials and scoped permissions. The blog says agents in private preview through the Frontier program can use their own approved credentials, with permissions scoped to the work they're given.
- Business process memory and triggers. Memory lets the agent keep relevant context over time. Triggers let it start work when an event happens or on a schedule.
Microsoft's example is a finance agent that prepares a weekly budget review. It collects the latest numbers, flags discrepancies, asks budget owners to explain them, folds in their replies as they arrive, and keeps a list of open questions before the meeting. That's an illustration, not a customer deployment or a measured result. Microsoft says people still make the important decisions.
The timing is a little confusing. This is not the first time these features have been announced. Several outlets reported that Microsoft announced three additions for this kind of work on October 15, 2026. They're agent accounts, business process memory, and triggers for agents built on the GitHub Copilot harness. They're rolling out through the Frontier program. So the November 4 post is a monthly summary of capabilities announced in October. It labels them a private preview through Frontier. Nothing is generally available.
Section summary: Persistent agents combine an identity, memory, and triggers. They're a Frontier preview for now, and Microsoft's finance example is a scenario, not a case study.
How agent accounts actually get turned on
The blog post doesn't explain setup, but Microsoft Learn now has a page called "Give your agent an account (Frontier)." It shows that admins have to do some work before any maker can start. Before users can create agents with accounts, admins must complete both steps in the following sections, in order.
- Activate the agent template. In the Microsoft 365 admin center, a tenant admin goes to All Agents, filters by Agent template: Yes, selects Copilot Studio (Frontier), and then selects Activate. In the activation wizard, at the Users step, select the users or groups that can create agents with accounts, and complete activation.
- Turn on the environment setting. In the Power Platform admin center, under Copilot Studio, select Agent accounts and turn on the setting for your environment. This setting lets you create user accounts and assign them to agents in the environment.
- Create the agent. The maker opens Copilot Studio Preview, creates a new agent, and selects Agent account under the agent's name.
What success looks like: the Agent account button shows up. If you turn it off, the Agent account button isn't available for agents in the environment. The most common problem: the button is missing. Microsoft's troubleshooting note is short: If Agent account isn't available, first verify that you're using Copilot Studio Preview and that your environment uses the early release cycle. Also note that your agent's account isn't active until you publish the agent. It uses a Microsoft 365 Frontier for AI Teammates license.
The fine print admins should read
The licensing costs more than it might seem at first. The consultancy 2toLead summarizes it: an agent account carries real cost: the maker's E5 plus Agent 365 or E7 license, one Microsoft 365 Frontier for AI Teammates license per agent, Copilot Credits for usage and the admin time to run a preview. The same analysis notes that agent accounts are used by agents powered by the GitHub Copilot harness, with usage-based billing in Copilot Credits for building, testing and running. Microsoft has since renamed that harness the Microsoft Copilot harness, which explains why both names show up in the documentation.
Memory is the bigger governance issue. Independent coverage of the preview says that an admin must turn off agent accounts altogether to turn off shared memory, and end users can't opt out. Another analysis adds that agents with accounts can't read documents protected by sensitivity labels. These are third-party readings of preview documentation, so check them against your own tenant. If they hold, a persistent agent that many people use could carry what it learned from one user into a conversation with another. Your data owners should know about that before the pilot, not after it.
Keep two identity features separate. New Copilot Studio agents get a Microsoft Entra Agent ID automatically. Full agent accounts with a mailbox and Teams presence need Frontier enrollment and extra licensing. One is an automatic identity for governance. The other is a licensed user-style account for working alongside people. They're not the same thing.
Persistent memory is also not brand new. Microsoft Learn's What's New page listed memory for harness agents in June 2026, which stores user preferences per user. The new "business process memory" is a separate announcement, and Microsoft hasn't said the two work the same way.
Section summary: Expect prerequisites to stack up: Frontier enrollment, an early-release environment, maker and per-agent licensing, and metered credits. Sort out your memory and sensitivity-label policies before a pilot.
Building agents: Cowork skills, MCP server, and in-chat connectors
Copilot Studio skills in Cowork are available through Frontier to users with a Microsoft 365 license. A maker describes a business need, drawing on existing files and processes, and Cowork helps build agents, apps, and workflows. Microsoft's example is an HR specialist with no automation background who builds an onboarding solution: an app that tracks new-hire readiness, approval workflows, and an agent that handles exceptions. Microsoft says Copilot Studio handles testing, governance, and deployment underneath, so admins keep visibility.
The Copilot Studio MCP server brings Copilot Studio into supported command-line interfaces and agent tools, including GitHub Copilot, Claude Code, and OpenAI's Codex. Developers can find resources, create and configure agents and workflows, run tests and evaluations, publish, and troubleshoot in plain language, all under existing organizational permissions. Microsoft is letting competing coding agents drive its own platform, which shows how much the MCP standard has caught on.
Connector setup in the conversation means makers no longer have to switch to a separate portal to set up connections. The guided flow reuses eligible existing connections and then returns the maker to the chat. Microsoft says this is generally available for Microsoft Copilot harness agents and for custom web agents on the Standard harness.
Customer conversations, A2A, and Work IQ
- Customer conversation harness (preview): a dedicated harness for customer-facing voice and chat agents. It brings together business knowledge, tools, specialist sub-agents (for billing, account support, or troubleshooting, for example), and escalation to human representatives. Testing, evaluation, deployment, and monitoring are built in.
- Agent2Agent support: the Microsoft Copilot harness can connect to agents built in Copilot Studio or in any framework that supports the open A2A protocol. It discovers what those agents can do and routes requests based on each agent's description, the user's request, and the conversation context. In practice, a well-written agent description now affects routing, so treat it like configuration rather than marketing copy.
- Work IQ (generally available): a context layer that pulls in email, meetings, files, chats, and organizational relationships. Microsoft Learn adds useful details: it's added by default to new agents on the newer harness, every workload is selected by default, it uses Copilot Credits, and write operations are on by default, subject to existing permissions. Admins who want least privilege should remove the workloads an agent doesn't need.
Workflows and cost controls
The blog lists three workflow features as generally available:
| Feature | What it does |
|---|---|
| Invoke Copilot capabilities | Calls Copilot Chat or Cowork from inside a workflow |
| Extract node | Pulls structured data from documents, emails, and images for later steps |
| Co-authoring | Lets several authors share and maintain one workflow |
Microsoft Learn's What's New page listed the Extract node and a preview Copilot node under September 2026, so "October GA" is how the blog frames it. The timeline isn't fully clear from Microsoft's own pages.
On costs, makers can now see how many Copilot Credits they use while building and testing, plus reports for each workflow run and each node. For admins, a preview Copilot Credit cost management experience in the Microsoft 365 admin center tracks usage, trends, and remaining capacity across agents, workflows, apps, Cowork, and Work IQ APIs. Admins can set user-based limits at the tenant or group level and agent-based limits at the environment level, configure alerts, and choose capacity packs, the Pre-Purchase Plan (P3), or pay-as-you-go billing. Microsoft didn't publish prices or credit amounts.
These controls matter more because of persistent agents. An agent that wakes up on a schedule and works with no one watching will keep spending credits even when nobody is using it. Microsoft's own harness guidance warns that the more capable runtime can overcomplicate simple jobs and raise latency and credit use without improving results.
The bottom line
This is a useful update, but it's still early. The persistent-agent package goes after a real gap: agents that act under their own identity, remember where a process stands, and pick work back up on their own. But it's a Frontier preview with no published GA date, no public list of supported triggers, no documented memory-retention controls, and no measured customer results.
Our advice, based on general enterprise rollout practice rather than anything Microsoft prescribes:
- Confirm Frontier enrollment and an early-release environment before you promise anyone a pilot.
- Keep the Copilot Studio (Frontier) template limited to a short, named list of makers.
- Set credit spending limits and alerts before the first triggered agent goes live.
- Review the shared-memory behavior with your compliance team, and limit Work IQ workloads to what each agent needs.
- Keep predictable approvals and record updates in deterministic workflows, and let the agent handle the exceptions.
Is your organization ready to give an AI agent a mailbox, a Teams presence, and a standing assignment? Copilot Studio can technically support that now, at least in preview. Whether your governance can keep up is a separate question.