What Microsoft announced
Microsoft used its Ignite 2026 conference, which opened on 17 November, to extend Agent 365 from cloud agents down to the AI agents running on staff laptops. These include coding agents such as Claude Code and GitHub Copilot CLI, along with the MCP servers that connect agents to files, databases and SaaS tools. Until now, IT teams on the Microsoft stack could govern agents built in Copilot Studio or Foundry. The agents a developer or analyst installed from a terminal mostly went unseen.
Microsoft says organisations have registered more than 50 million agents in Agent 365 since the product launched in May 2026. The new local agent features break down as follows, according to Microsoft's Agent 365 team:
- Defender's local agent discovery is generally available on managed Windows and macOS devices. It links each agent to a device and a user. Defender and Global Secure Access together can identify more than 60 local AI agents and the MCP servers they're configured to use.
- Shadow AI discovery, which adds unsanctioned agents to the Agent 365 registry, is in preview.
- Entra Agent ID for local agents is in preview. It starts with GitHub Copilot CLI and Claude Code and gives each agent an identity separate from the person running it.
- Defender runtime protection for local agents is generally available on Windows. It inspects the agent loop and can block malicious instructions in audit or block mode.
- Purview data loss prevention for agents including Claude Code, Claude Desktop and GitHub Copilot CLI is in preview.
The Intune compliance policy is the part with teeth
For most IT teams, the Intune change has the biggest practical effect. WindowsForum's analysis of the Intune announcements explains that a new software compliance policy, now in public preview, can mark a device as noncompliant when a specified unsanctioned agent is installed. Conditional Access then blocks the device from corporate resources until the user removes the agent. The policy doesn't uninstall anything. It cuts off access through the same compliance chain many organisations already use for OS versions and disk encryption, a comparison Microsoft MVP Jeroen Burgerhout draws in his reading of the Book of News.
The targeting is narrower than the announcement suggests. According to WindowsForum, the policy identifies software by MSI product code, and an update on 18 November added Store apps by package full name. npm packages, editor extensions and standalone executables can't be selected. Many developer agents are installed in exactly those ways, so a policy aimed at "Claude Code" may miss the copy a developer pulled down through npm. Compliance is also checked against the host application, so an agent with several host apps may need several entries. A wrong match can lock someone out of email and files when the agent isn't even running.
Burgerhout adds that Microsoft hasn't explained how discovery decides what counts as an agent, and that licensing is rarely stated for the Ignite items. Most of the features are preview or private preview, and he treats them as direction rather than a rollout plan.
What it costs
Microsoft's licensing notes say core usage insights for Microsoft's own agents need only a qualifying Microsoft 365 subscription. Managing non-Microsoft agents and the advanced governance and security features require an Agent 365 or ME7 licence. When Microsoft announced the products in March, The Register reported Agent 365 at US$15 per user per month and Microsoft 365 E7 at US$99. Microsoft hasn't published which of the new local agent controls each licence includes, and NZ dollar pricing isn't in the announcements. Ask your licensing partner before you budget.
A new Agent 365 Cost Management tool, in preview according to the Book of News, lets admins set spending limits and alerts on usage-based AI services. The Agent 365 team says the tool tracks spend by user, agent and model. That matters for firms where AI agents are billed per token or per action and nobody owns the total.
Why it matters for NZ businesses
Local agents are the hardest kind for a mid-sized NZ firm to see. A developer at a 60-person software house or an analyst at a distributor can install a coding agent in minutes. That agent then reads the local repository, runs shell commands and calls whatever MCP servers it's been pointed at, all under the user's own credentials. In logs, the agent's activity looks the same as the user's. Microsoft's move to give local agents their own Entra identity addresses that problem directly, but the identity feature is in preview and covers two agents so far.
For organisations already on Intune and Conditional Access, the new controls fit into a process the IT team already runs. That lowers the effort of getting an inventory compared with buying a separate agent security product. The limits are real, though. Discovery covers managed devices only, so contractors' laptops and unenrolled personal machines stay invisible. The compliance policy also can't target the common npm and executable installs. A firm that turns on blocking without first checking what's installed is likely to cut off its most productive staff while missing the agents it worried about.
None of the Microsoft tooling decides which agents your business should allow. A block list needs an approved list behind it, along with a named owner for each approved agent and a rule for which systems each one may touch. That governance work is where we see NZ organisations stall, and our AI consultancy work starts there before any tool gets switched on.
Practical next steps
- Run Defender's local agent discovery, which is generally available, on managed Windows and macOS devices and see what's already installed before writing any policy.
- Compare the discovered agents with the tools staff have told you they use. The gap shows how much shadow agent use you have.
- Write a short approved-agent list. For each agent, name an owner and list the repositories, data and MCP servers it may reach.
- Pilot the Intune software compliance policy on a small IT group in report mode first. Confirm the MSI or Store targeting matches how your staff installed each agent.
- Start Defender runtime protection in audit mode, as Microsoft recommends for approved CLI agents, and review alerts for a few weeks before you switch on blocking.
- Ask your Microsoft partner which features need Agent 365 or E7 licences, what the NZD cost is, and whether preview features carry support commitments.
- Set spend alerts on any usage-billed AI service now, even if you aren't ready for Agent 365 Cost Management.
If you're planning agent deployments beyond coding assistants, our AI agents team can help you design the identities, permissions and approval steps before the agents reach production data.
Sources
- Microsoft Ignite 2026 Book of News
- What's new in Microsoft Agent 365: Ignite 2026 (Microsoft Tech Community)
- Microsoft Ignite 2026: your guide to Agent 365 (Microsoft Tech Community)
- Microsoft Intune Ignite 2026: AI agent governance, Windows Server management and preview caveats (WindowsForum)
- Microsoft Ignite 2026: the Intune bits of the Book of News (Jeroen Burgerhout)
- Microsoft 365 confirms new premium tier, stuffed with AI and few discounts (The Register)
- AI Daily Digest November 18, 2026 (DEV Community)
Want to know what this means for your organisation? Our team in Auckland can show you your top three opportunities in a free 30 minute assessment.
Get my free assessment How we help