Field SOP
Field SOP

Microsoft MXC GA: OS-Enforced Containment for Agents

Putting agent permissions behind the operating system is the idea behind Microsoft Execution Containers (MXC). Developers declare the files and network destinations a workload needs; MXC enforces those policies at runtime, keeping them outside the workload. This SOP walks through the boundary principle, the Containment/Identity/Manageability pillars, and the official split in supported-agent counts.

Published November 18, 20269 min read
<!-- microsoft-mxc-agent-sandbox-sop | sop | Microsoft MXC GA: OS-Enforced Containment for Agents -->

Putting agent permissions behind the operating system is the idea behind Microsoft Execution Containers, called Microsoft MXC here. Developers declare which files and network destinations a workload needs; MXC uses the appropriate container to enforce those policies at runtime, and the policy stays outside the workload so an agent and the code it generates cannot grant themselves more access. The title's OS-Enforced Containment for Agents names that operating-system line. This is a governance and operations SOP focused on five steps and boundary verification, not a product launch recap; where official material gives no command, configuration item, or path, none is invented here. The distinction matters because the policy is declared outside the agent, so neither the agent nor its generated code can widen its own permissions, a point the launch blog makes explicitly. As of November 17, 2026, the first-party Microsoft GA record begins on October 7 in the Windows Developer Blog with the SDK v1.0.0 first release, reaffirmed on November 5 in the Security Blog; a secondhand brief anchors the GA to a November 16 security day, while Microsoft marks Ignite 2026 as November 17 to November 20, and no standalone November 16 GA announcement was found. The two accounts differ and are presented side by side, without picking one.

First, see why the control point cannot live inside the agent

Start with one principle: an agent cannot be its own security authority. The launch blog says "An agent cannot be its own security authority"; the October 19 security blog sharpens it to "the model cannot be its own security boundary"; another official post says "Model transparency will not replace containment." None is a verbatim claim that prompts or allowlists are not security controls, but all make the point that a boundary relying on model self-restraint lets the measured party hold the ruler. The control point belongs outside the workload, enforced by an independent layer.

The official definition calls MXC a policy-driven execution layer for untrusted or dynamically generated workloads. Model output, plugins, tools, the agent harness, or the whole agent can be contained; developers declare required resources and MXC enforces the boundary with the right container. The launch blog names three pillars: Containment limits what an agent can access and do; Identity distinguishes an agent's activity from a person's; Manageability gives organizations tools to govern access and monitor activity. The November 5 update calls it an operating system-enforced containment layer with least-privilege controls across files, networks, processes, and user interactions. A GA label without that OS-level force would be only a marketing line.

Step 1, enable: from installing the SDK to seeing who can connect

The first step is enablement, and official material stops at a clear boundary: install the SDK through a package manager (Rust, .NET, Node), or, without an SDK, use platform executor binaries that accept JSON container-creation requests against the stable schema, meant for testing or cases where the SDK cannot be embedded in an app. Agent 365 needs at least one qualifying user license to enable, and works best with Microsoft E5 as a prerequisite. No enterprise one-click onboarding command, configuration item, or path is published, so none is invented here. Managing MXC process containers with Intune on Windows 11 is described only as "soon" (launch blog) and "upcoming" (November 5 update), with no date or procedure; any fixed date is unreliable.

The shipped tier must stay strict. Official wording "already support" covers GitHub Copilot, OpenAI Codex, and Replit, plus OpenClaw, LM Studio, and Unsloth AI, with the Copilot, Codex, and Replit scenarios singled out in the official post as demonstrating a demanding developer workflow. The "will be releasing support" tier includes Anthropic Claude Code, Box, Egnyte, Heidi Health, Nous Research's Hermes Agent, Manus, Perplexity, Raycast, and Simular. NVIDIA is a separate line: "NVIDIA has integrated OpenShell into MXC," adding file and inference-service policy controls, advanced networking, credential management, and enterprise OCSF auditing. Claude Code is coming soon, not shipped; calling it available is a common misreading.

Step 2, choose a backend: pick the lightest that fits the blast radius

The second step picks one of the four backends by blast radius, taking the lightest that satisfies the isolation requirement. These four are the product-level categories in the launch blog; the repository lists more backends, including Windows Sandbox, Hyperlight, Bubblewrap, LXC, and Seatbelt. The four official categories:

Each backend carries distinct security properties, and the guidance says workloads should be evaluated for fit rather than assigned one by default. The process container runs on Windows 11, macOS, and Linux for lightweight, responsive workloads including generated code and tool execution, using AppContainer on Windows, Seatbelt on macOS, Bubblewrap on Linux. The session container is Windows 11 only, for long-running agents needing a desktop or stronger separation from the interactive user; it runs under a distinct Windows account and session, separating the agent's desktop, clipboard, UI, input, and active session. Note the official caveat: session isolation separates user and session but does not restrict what the workload can access; on Windows it refuses file-system and UI policies and requires egress, ingress, and host loopback set to allow, so workloads needing those controls require another backend. The WSL container (WSLc), also Windows 11 only, serves Linux-first toolchains through WSL. The experimental MicroVM, marked experimental, runs on Windows 11 and Linux for higher-risk workloads, offering hardware-enforced isolation and full Linux workload compatibility.

MicroVM needs explicit opt-in, cannot be created through the V1 SDK API, and is currently Windows x64 only, so calling it routinely available is inaccurate. On Windows, process isolation is described as common because it suits workloads that need part of the host with low startup cost and task-scoped privileges. Selection guidance: check host OS and architecture, required guest environment, file, network, and UI access, whether commands share state, and limits on startup time, resources, deployment, and interaction, then judge whether a backend can enforce every needed control. Runtime discovery on .NET uses GetPlatformSupport and GetAvailableBackends, and on Windows MxcContainer.Probe checks compatibility without creating a container; the docs stress discovery does not validate the full policy.

Step 3, verify the blast radius: test both positive and negative

The third step verifies the boundary. The official acceptance line: run positive and negative tests together, since a successful start only shows MXC accepted the configuration; a negative test such as writing outside a granted path shows whether restrictions actually hold. Passing only the positive case verifies nothing.

Three modes help. Enforcement blocks and records, and is the production mode. Learning blocks and records, for diagnosing failures and checking the policy grants only required access. Permissive allows and records, observing activity without enforcing; it does not bypass other OS or organizational restrictions, but offers no real protection and should not remain in production, a deployment recommendation rather than a verbatim official conclusion. For Windows ProcessContainer workloads, a learning mode aids policy authoring. Defender-side exposure mapping can be added, with Kusto Query Language for advanced hunting over discovery data. The exposure map shows visual relationships among local agents, devices, identities, and the resources those identities can reach, helping assess potential impact based on the privileges of the user running an agent. No blast-radius score or threshold is published, so verification is qualitative.

One trap: the official line reads that sandbox teardown typically does not undo writes to granted host directories or operations through allowed network services. Sandbox teardown is not rollback; choose output locations and decide which results to keep for later tasks.

Step 4, configure Defender observation: two official numbers side by side

The fourth step adds observation. Official prerequisites: commercial cloud (sovereign and national clouds unsupported), Microsoft Defender for Endpoint Plan 2, devices onboarded to Defender for Endpoint, Windows antivirus platform 4.18.26090.9 (2609) or later and macOS Defender for Endpoint 101.26072.0017 or later. When met, agent discovery starts automatically with no extra deployment or scripts; for process-based discovery, seeing a process run can create an inventory entry, though discovery does not prove an AI service was used. Observation lives in the Defender portal under Assets, AI agents, with risk level, indicators, recommendations, and active alerts; viewing risk information needs Microsoft 365 E7 or Agent 365 with Defender for Endpoint Plan 2. Runtime protection monitors the agent loop and blocks malicious instructions before the actor acts.

On the count, two official posts disagree and both are reproduced. The October 19 blog says discovery expands by 35 additional types, bringing the total to 60 supported local agents, naming Claude Code, OpenAI Codex, and GitHub Copilot. The November 5 blog says Defender discovers more than 50 supported local agents. The earlier post has the larger number and product documentation gives no count, so the inconsistency is in official material; the 60 and 50 versions are presented side by side with no forced resolution.

Agent 365 also has two official wordings: the Learn root page calls it an observability platform for AI, while a pricing-related page title still reads the control plane for agents. Microsoft has not written verbatim that it covers agents, models, and assistants as three parallel categories, so that scope is not added here.

Step 5, do the math: per-user billing and token unit price in one ledger

The fifth step makes costs explicit. Two official billing statements need separating. Agent 365 is billed per user: an official news release gives 15 dollars per user per month, and Microsoft 365 E7, bundling Copilot and Agent 365, is 99 dollars per user per month; E7 is described as E5 plus Copilot plus Agent 365, which is why its per-user price is higher. MXC ships with the SDK; the captured README license section only says "see LICENSE.md for details" and does not confirm a license name. Token unit prices are outside Microsoft's MXC or Agent 365 documentation, belonging to model vendors. Official material describes Agent 365 as a single platform IT and security leaders can use to observe, manage, govern, and protect agents across the organization, generally available to commercial users from May 1.

"Same ledger" means both land on one cost sheet: the per-user Agent 365 seat fee and the token unit price that scales with usage. Microsoft has published no unified ledger or conversion rule merging them, so the two official statements are listed side by side and no formula or unit price is fabricated. For an enterprise that means two cost lines to reconcile, a per-user seat fee and a usage-based token charge governed by different vendors' rate cards. The math loops back to a trap: Intune, Entra, and Agent 365 integration for enterprise governance is "upcoming" in the November 5 wording, while MXC itself is GA; do not read "MXC GA" as "all management integration is GA too." To follow token costs, see the Token Cost SOP.

FAQ

Q1: Is MXC GA dated October 7 or November 16? A1: As of November 17, 2026, the first-party Microsoft GA record begins on October 7 in the Windows Developer Blog, reaffirmed on November 5 in the Security Blog. November 16 appears only in a secondhand brief; Microsoft marks Ignite 2026 as November 17 to November 20, and no standalone November 16 GA announcement was found. The two accounts are presented side by side.

Q2: Can Claude Code connect to MXC now? A2: Per the official launch blog, Claude Code is in the "will release support" tier, not shipped; the shipped tier includes GitHub Copilot, OpenAI Codex, and Replit. Calling "coming soon" "already available" is a common misreading.

Q3: Does Defender find 60 or more than 50 local agents? A3: Official posts disagree: October 19 says 35 additional types for a total of 60, November 5 says more than 50. Both the 60 and 50 versions are reproduced, the inconsistency noted, with no forced choice.

Q4: After sandbox teardown, do files and operations roll back? A4: Not automatically. The official line says sandbox teardown typically does not undo writes to granted host directories or operations through allowed network services. Choose output locations carefully and decide which results to keep; sandbox teardown is not rollback.

Q5: When will Intune manage MXC process containers on Windows 11? A5: Official wording is only "soon" and "upcoming," with no date or procedure. Any fixed date or invented command is unreliable; follow the SDK's built-in policy until Microsoft publishes steps.

Related reading

This article is AI-assisted and human-edited. Last updated: 2026-11-18

FAQ

Is MXC GA dated October 7 or November 16?
As of November 17, 2026, the first-party Microsoft GA record begins on October 7 in the Windows Developer Blog, reaffirmed on November 5 in the Security Blog. November 16 appears only in a secondhand brief; Microsoft marks Ignite 2026 as November 17 to November 20, and no standalone November 16 GA announcement was found. The two accounts are presented side by side.
Can Claude Code connect to MXC now?
Per the official launch blog, Claude Code is in the "will release support" tier, not shipped; the shipped tier includes GitHub Copilot, OpenAI Codex, and Replit. Calling "coming soon" "already available" is a common misreading.
Does Defender find 60 or more than 50 local agents?
Official posts disagree: October 19 says 35 additional types for a total of 60, November 5 says more than 50. Both the 60 and 50 versions are reproduced, the inconsistency noted, with no forced choice.
After sandbox teardown, do files and operations roll back?
Not automatically. The official line says sandbox teardown typically does not undo writes to granted host directories or operations through allowed network services. Choose output locations carefully and decide which results to keep; sandbox teardown is not rollback.
When will Intune manage MXC process containers on Windows 11?
Official wording is only "soon" and "upcoming," with no date or procedure. Any fixed date or invented command is unreliable; follow the SDK's built-in policy until Microsoft publishes steps.

Related

Field SOP

GitHub Copilot: Six Models Deprecated December 14, What to Fix

Six GitHub Copilot models are deprecated on December 14, 2026, and the official November 13, 2026 changelog gives six migration sentences but no configuration file name, model ID string or migration script. This SOP turns those sentences into a checkable list: replace hardcoded model names in workflows and integrations, enable model policies in Copilot settings, and understand what global default, the model selector, Copilot Chat, inline edits and code completions each do. All six model names and suggested alternatives are quoted as published, Kimi K3 stays To be announced, and the Claude Haiku 4.5 annual-plan exception for individual subscribers is spelled out.

Nov 16, 20269 min read
Field SOP

Bedrock AgentCore V2 Snapshots: Four Patterns, One Pre-warm

Turns the four official AWS section names into a checkable sequence for Bedrock AgentCore runtime V2 session startup: measure Latency and OverheadLatency in CloudWatch first, move imports and static data into init, use Working with hooks for build-time large files, then send a background warm-up request with Pre-warm on user engagement and keep the session ID. It records the 120-second /ping gate, that the V2 default applies only to new and updated runtimes, that existing ones need explicit platformVersion=V2, and that the sample code is for non-production use. The Region and cold-start discrepancies between the two official pages are kept side by side.

Nov 15, 20269 min read