WED, NOVEMBER 18, 2026
Independent · In‑Depth · Practitioner‑Tested
AI Agents

Microsoft's Four Agent Control Layers Compared (2026): Which One Actually Stops What

Entra Agent ID, the MCP Firewall, Intune device governance and Windows process containers all shipped at Ignite 2026 and all claim to control agents. They stop four different things, and only two are generally available.

🕐 7 min read 👁 66 views 📅 Nov 18, 2026
Four layers, four different jobs
● Identity Entra Agent ID. Tells you which agent did what. Preview.
● Network Global Secure Access and the MCP Firewall. Controls what an agent can reach. Partly GA, firewall in preview.
● Device Intune with Agent 365. Keeps a device carrying an unapproved agent away from company data. Preview.
● Runtime Windows process containers. Limits what a running agent can touch on the machine. Preview, learning mode GA.

Microsoft announced agent controls at four layers at Ignite 2026. Coverage has tended to treat them as one announcement. They solve different problems and they fail in different ways, so it is worth separating them before deciding what to turn on.

Layer 1 - Entra Agent ID: attribution

Entra automatically assigns an Agent ID to local agents it discovers on managed devices, starting with GitHub Copilot CLI and Claude Code. It also extends to Microsoft cloud agents, beginning with Autopilot agents.

What it stops: nothing, on its own. What it produces is attribution - agent activity separated from human activity in the audit trail. That sounds modest until you try to investigate an incident where an agent and its operator share one identity, at which point every other control you own is reporting against the wrong subject.

Status: preview, both variants.

Layer 2 - MCP Firewall and Global Secure Access: reach

Global Secure Access unifies Internet Access and Private Access across apps, SaaS, AI and agents. Three pieces matter here. Agentic traffic acquisition selectively captures agent traffic for visibility and enforcement, and is generally available. Network visibility and control over local agents, their destinations and MCP servers is in preview. The MCP Firewall controls access to MCP servers, tools, resources, prompts, protocols and transports, and is in preview.

What it stops: an agent reaching a server, a tool or a prompt it should not. The MCP Firewall is the only control in this set with granularity below the server - it can govern individual tools and prompts, not just whether the connection opens.

Status: agentic traffic acquisition GA; local agent control, MCP Firewall, unsanctioned agents and adaptive access all preview.

Layer 3 - Intune and Agent 365: device posture

Intune discovers agents on managed devices through Agent 365, marks a device carrying an unapproved agent as noncompliant, and Conditional Access then blocks that device from company data. Approved agents run inside Microsoft Execution Containers under Intune policy.

What it stops: the whole device, which is the bluntest and sometimes the only useful instrument. It is the only layer here that can deny access to company data outright on the basis of what is installed.

Status: preview.

Layer 4 - Windows process containers: runtime

Process container controls isolate agents with COM and IPC governance, granular network policies and identity protections. A learning and permissive mode lets teams observe agent behaviour before enforcing policy.

What it stops: a running agent reaching other processes, other applications or the network in ways policy forbids. This is the layer that matters when the agent executes code you did not write.

Status: controls in preview; learning and permissive mode generally available.

The order that makes sense

Identity first, because the other three generate evidence you cannot attribute without it, and because the learning modes are only informative once you know which agent is which. Device governance second if your actual worry is agents arriving on laptops without approval. Runtime containers where agents execute code. The MCP Firewall last, not because it is least important - it may end up the most important of the four - but because it is the newest and the only one with no generally available equivalent to fall back on.

One practical note: the container learning and permissive mode being generally available while the enforcement controls are in preview is the right way round, and it is worth using. Watch what your agents do for a few weeks before writing policy about it.

FAQ

Do I need all four?
They address different failures, so most organisations running agents at scale will end up with at least identity and one enforcement layer. Identity is the one with no substitute.
Which of these can I deploy in production today?
Microsoft MCP Server for Enterprise, agentic traffic acquisition, and the process container learning and permissive mode are generally available. Agent ID, the MCP Firewall, Intune agent governance and the container enforcement controls are all preview.
Does any of this work with non-Microsoft agents?
Yes. Entra Agent ID names GitHub Copilot CLI and Claude Code as its first targets, and Intune's discovery is based on what is found on the device rather than on vendor.
What is the MCP Firewall for specifically?
Controlling which MCP servers, tools, resources, prompts, protocols and transports an agent may use. It is the only layer here that reaches individual tools and prompts.
⚖ Our Verdict

Layers, not options. Start with Entra Agent ID: without identities the other three produce logs you cannot attribute. Add Intune device governance next if unapproved agents on laptops are the worry - it is the only layer that can deny a device company data outright. Watch the MCP Firewall rather than deploy it. Only three of these are generally available; the rest are preview.