Microsoft just answered a question most enterprises have been quietly wrestling with: where do AI agents actually run? Not the model inference part, that is handled by Azure OpenAI or whatever LLM provider you pick. The other part. The part where an agent needs to open Excel, fill out a form in SAP, click through a legacy web app, or interact with software that has no API. Windows 365 for Agents provisions dedicated cloud PCs for these workloads at $0.40 per hour, managed through the same Intune and Entra ID infrastructure enterprises already use for human employees.
This is not a VM you spin up and SSH into. It is a full Windows desktop environment that computer-use agents interact with visually, reading screens, clicking buttons, and typing into fields the same way a human contractor sitting at a desk would.
How Computer-Use Agents Actually Work on Windows 365
The agents running on Windows 365 for Agents are not traditional RPA bots with hardcoded selectors. They are computer-use agents (CUAs) that interpret screen content using AI vision models, reason about what they see, and decide what to do next. Think of it as giving an AI a remote desktop session and saying “process this invoice.”
The agent takes a screenshot of the current screen state. A vision model (typically GPT-4o or similar) interprets what is on screen: buttons, text fields, dropdown menus, error messages. The model generates an action plan: click this button, type this text, scroll down, wait for this element. The action executes, a new screenshot is taken, and the loop repeats.
Why This Matters More Than API Integration
The obvious question: why not just build API integrations instead? Three reasons.
First, most enterprise software does not have comprehensive APIs. SAP’s GUI, legacy internal tools built in the 2000s, government portals, insurance claim systems: these applications were built for humans sitting at screens. Rebuilding them with API access would take years and cost millions. A computer-use agent can interact with them tomorrow.
Second, CUAs adapt when UIs change. A traditional RPA bot breaks when a button moves three pixels to the right or a form field gets renamed. A vision-based agent processes what it sees semantically. It understands “Submit” and “Send” are the same intent, even if the label changes between software versions.
Third, compliance. Many regulated processes require that actions happen through the same interface a human auditor can review. Running an agent through the GUI produces the same audit trail as a human operator, which is easier to explain to regulators than a series of API calls.
The Check-In/Check-Out Architecture
Windows 365 for Agents uses a pool-based model that borrows concepts from how enterprises manage shared workstations. IT administrators create host pools of cloud PCs, pre-configured with the right applications and security policies.
When an agent needs to do work, it checks out a cloud PC from the pool. The agent gets an isolated Windows environment with its own session, its own file system, and its own application state. When the task finishes, the agent checks the cloud PC back in, and the environment resets for the next agent to use.
Two pool types handle different latency requirements:
Warm pools keep cloud PCs pre-provisioned and ready. An agent can start working within seconds. This is the right choice for customer-facing workflows where response time matters, like processing a support ticket that requires looking up information across multiple legacy systems.
Cold pools provision cloud PCs on demand. Startup takes longer (minutes rather than seconds), but you only pay when agents are actively working. Batch processing jobs that run overnight, like reconciling data across systems, fit this model.
Pricing Realities
At $0.40 per hour (rounded up to the next full hour), a single agent task that takes 15 minutes costs $0.40. An agent that runs continuously for 8 hours costs $3.20 per day, or about $70 per month for 22 working days.
Compare that to a human contractor doing the same repetitive GUI work: $25-50 per hour in the US, $15-25 in offshore locations. The agent is roughly 30-60x cheaper per hour, assuming it can handle the task reliably.
But the rounding matters. If your agent workflow involves many short tasks (2-3 minutes each), each one gets billed as a full hour. Batching tasks into longer sessions becomes an optimization problem that did not exist before.
Agent 365: The Control Plane
Windows 365 for Agents provides the compute. Agent 365, going generally available on May 1, 2026, provides the governance layer. Think of it as the “HR department” for your AI agents: it handles identity, permissions, monitoring, and lifecycle management.
Every agent gets an Agent ID in Microsoft Entra ID. This means agents are subject to the same conditional access policies, least-privilege models, and zero-trust principles as human users. An agent processing expense reports can be restricted to only the finance applications it needs, during business hours, from specific network locations.
What Agent 365 Manages
Discovery and catalog. A central registry where enterprises can see all deployed agents, whether built by Microsoft, third-party partners, or internal teams. No more shadow agents that IT does not know about.
Governance policies. Administrators define what agents can access, what data they can touch, and what actions require human approval. These policies follow the agent across different execution environments, including Windows 365 cloud PCs.
Monitoring and audit. Real-time visibility into what agents are doing, with screenshot-based audit trails for cloud PC sessions. If an agent does something unexpected, the audit trail shows exactly what it saw and what it clicked.
Lifecycle management. Deploying, updating, and retiring agents through the same change management processes enterprises use for other software. Agent versioning, rollback capabilities, and staged rollouts are built in.
The M365 E7 Play: Agents as Licensed Employees
Microsoft is also introducing Microsoft 365 E7, a $99/user/month bundle launching May 1 alongside Agent 365. It packages M365 E5 ($60), Microsoft 365 Copilot ($30), Entra Suite ($12), and Agent 365 ($15) into a single SKU. Buying separately would cost $117, so the bundle saves about 15%.
The strategic signal is louder than the pricing math. Microsoft is positioning agents as entities that need licenses the same way employees do. An agent that processes invoices needs access to Outlook (to receive invoices via email), SharePoint (to store documents), Teams (to notify humans of exceptions), and Entra ID (for identity). That stack of entitlements looks a lot like a human employee’s license.
This creates an interesting dynamic for IT budgeting. Today, enterprises plan headcount and license counts together. Tomorrow, they will need to plan agent counts alongside both. A department might have 50 human employees and 200 licensed agents, each with different permission profiles but all consuming Microsoft 365 resources.
Who Is Already Building on Windows 365 for Agents
Microsoft did not launch this in a vacuum. Several agent-building companies are already part of the program:
Manus AI is using Windows 365 cloud PCs to power intelligent PowerPoint creation and editing workflows. Their agents open PowerPoint on the cloud PC, manipulate slides based on natural language instructions, and save the results. This works with the full PowerPoint desktop application, not just the limited API surface.
Fellou builds an agentic browser that merges AI with daily web workflows. Windows 365 gives Fellou’s agents an isolated browser environment where they can navigate websites, fill out forms, and extract information without touching the user’s local machine.
Genspark focuses on marketing content and slide generation. Their agents use the cloud PC environment to work with design tools and content management systems that only have GUI interfaces.
Simular, which raised a $21.5 million Series A, was one of the early partners. Their agent technology specializes in controlling desktop applications through visual understanding.
What is notable about these partners: none of them are building traditional chatbots or API-only agents. They all need a real desktop environment to do their work. That is the niche Windows 365 for Agents fills.
What Enterprise Teams Should Evaluate Now
If you are running or planning AI agent workloads in an enterprise, three things to figure out before Agent 365 goes GA on May 1:
Inventory your GUI-dependent processes. Which workflows currently require humans to click through applications? Which of those are repetitive enough that a computer-use agent could handle them? Start with the processes where you already have RPA bots, since those are proven candidates for GUI automation, and CUAs can handle the cases where RPA breaks.
Model the cost. The $0.40/hour pricing sounds cheap until you multiply it by hundreds of agents running thousands of tasks per day. Map out your expected task volumes, average task durations, and the hourly rounding impact. For many workloads, the math still favors agents over humans by a wide margin, but you want the real number, not the marketing number.
Plan the identity architecture. Agent 365 requires Entra ID integration. Each agent needs an identity, permission set, and governance policy. If your Entra tenant is already complex (and whose is not), adding hundreds of agent identities is a governance project, not just a technical one. Start talking to your identity team now.
Frequently Asked Questions
What is Windows 365 for Agents?
Windows 365 for Agents is a Microsoft service that provisions dedicated cloud PCs for AI agent workloads. Computer-use agents get their own virtual Windows desktop environments where they interact with applications visually, the same way a human would, using AI vision to read screens and perform actions. The service is managed through Intune and Entra ID and billed at $0.40 per hour.
How much does Windows 365 for Agents cost?
Windows 365 for Agents uses pay-as-you-go pricing at $0.40 per hour per cloud PC session. Usage is rounded up to the next full hour, so a 15-minute task costs $0.40. For continuous 8-hour workdays, that is about $3.20 per day or roughly $70 per month. The upcoming M365 E7 bundle at $99/user/month combines E5, Copilot, Entra Suite, and Agent 365 governance.
What is the difference between Windows 365 for Agents and regular RPA?
Traditional RPA bots use hardcoded selectors and scripts that break when UI elements change. Windows 365 for Agents runs computer-use agents that interpret screens visually using AI vision models. These agents understand UI elements semantically, adapting when buttons move, labels change, or layouts shift between software versions. They reason about what they see rather than following brittle scripts.
What is Microsoft Agent 365?
Agent 365 is Microsoft’s governance and control plane for AI agents, going generally available May 1, 2026. It provides agent identity management through Entra ID, policy enforcement, monitoring with audit trails, and lifecycle management. It lets enterprises discover, govern, and secure all their agents, whether built by Microsoft, partners, or internal teams, using the same security infrastructure they use for employees.
Which companies are building on Windows 365 for Agents?
Early partners include Manus AI (intelligent PowerPoint creation), Fellou (agentic browser for web workflows), Genspark (marketing content and slide generation), and Simular (desktop application control through visual understanding). All of these companies build computer-use agents that need real desktop environments rather than just API access.