Microsoft Copilot Shifts to Agent Governance: Claude Checks GPT Work, Screen Agents Go Live

Microsoft has repositioned Copilot from a chat assistant into a deployment and governance platform for autonomous agents — a shift that took concrete, purchasable form on May 1, when the company made Microsoft 365 E7 and Agent 365 generally available. The move follows a wave of Wave 3 announcements beginning March 9 that, taken together, redrew the boundaries of what enterprise AI infrastructure means: a licensing tier that governs agent fleets, a feature that lets agents drive software through a screen without an application programming interface, and a research workflow in which Anthropic's Claude audits every response OpenAI's GPT produces before the user sees it.

Enterprise IT teams now face a specific decision: whether $99 per user per month for E7 — or $15 for Agent 365 alone — represents a defensible investment in a governance layer they will need as agent deployments scale, or a premium for capabilities they are not yet ready to deploy at the autonomy levels that justify the price.

Microsoft 365 E7 Frontier Suite Bets on Agent Governance at $99 Per User

Microsoft 365 E7, which Microsoft calls the Frontier Suite, is the company's first new enterprise licensing tier since E5 launched in 2015. It bundles four components that many firms currently purchase separately: Microsoft 365 E5, Microsoft 365 Copilot, the Entra Suite, and Agent 365. Microsoft describes E7 as designed for a "human-led, agent-operated enterprise," and positions the $99 per user per month price as a discount versus buying the components separately; partner analyses have estimated the saving at between $14 and $30 per user depending on an organization's existing subscriptions. A Teams-excluded variant is priced at $90.45 per user per month.

The centerpiece is Agent 365, a control plane for governing autonomous agents at scale. As enterprises begin handing agents write-access to core systems, the governance problem becomes acute: who authorized that agent, what did it touch, and what did it decide? Agent 365 addresses this through Entra for agent identity, Defender for protection, and Purview for compliance. The admin center also adds a Shadow AI page to detect and block unauthorized agents — not only Microsoft's own but third-party tools, with detection extending to agents built outside the Microsoft ecosystem, including GitHub Copilot CLI and Claude Code.

Microsoft launched E7 with striking internal figures, though these are vendor-reported and have not been independently audited: within two months of preview, the company said tens of millions of agents had appeared in the Agent 365 registry, and that it has visibility into more than 500,000 agents operating across Microsoft's own internal environment, generating more than 65,000 employee responses a day. Analyst firm IDC, in a study commissioned by Microsoft, projected 1.3 billion active AI agents in enterprise environments by 2028. Whether or not those projections hold, the strategic intent is clear: Microsoft wants the registry where enterprise agents are governed to be its own, and the $99 bundle is how it embeds that registry into the purchasing decision.

An active Federal Trade Commission antitrust probe, ongoing as of this spring, is examining whether Microsoft's AI and cloud bundling practices — of the type that produced E7 — create anticompetitive lock-in. Microsoft has said it remains committed to fair competition. No enforcement action has resulted.

How Do Computer-Using Agents in Copilot Studio Work?

The second move closes a long-standing gap in enterprise automation. On May 13, Microsoft made computer-using agents generally available in Copilot Studio, rolling the feature to all commercial Power Platform geographies, with sovereign clouds excluded from the initial release.

These agents interact with websites and desktop applications through the user interface — clicking, typing, and reading the screen — rather than calling an application programming interface. The models available at general availability are OpenAI's computer-use architecture and Claude Sonnet 4.5; that distinction matters for enterprise procurement, as only generally available models carry contractual service-level guarantees.

That architecture matters because a significant share of enterprise work still runs through line-of-business applications with incomplete or nonexistent programmatic interfaces. Microsoft's argument is that traditional automation built on brittle scripts breaks the moment a vendor changes a screen layout, whereas a vision-and-reasoning agent can adapt to the changed interface without re-scripting. Graebel, a talent-mobility firm managing cross-border employee relocations for multinational clients, worked with Microsoft on the feature's design: Graebel's proprietary relocation platform had no automation interface, and earlier robotic-process-automation attempts proved too rigid for the variability of unstructured relocation request emails. Computer-using agents address that specific category of integration problem.

The security implications deserve naming. Tenable's research team demonstrated that a Copilot Studio agent could be prompted into disclosing customer credit card records and altering stored prices through a straightforward prompt injection, with no specialized hacking skills required. Separately, Capsule Security discovered CVE-2026-21520, a CVSS 7.5 indirect prompt-injection vulnerability in Copilot Studio that was patched in January 2026. Computer-using agents, which touch live systems and can take actions across previously separated application boundaries, carry those risks at elevated consequence. Agent 365's governance layer is Microsoft's answer; bundling these agents with governance controls is what makes them deployable rather than merely impressive, but governance makes the risk observable and manageable — it does not eliminate it.

GPT Drafts, Claude Critiques: Multi-Model Research Explained

The third move is the most conceptually distinctive. On March 30, Microsoft announced Critique, a feature for the Researcher agent in Microsoft 365 Copilot. The design is one-directional: OpenAI's GPT drafts a response to a research query, and Anthropic's Claude reviews it for accuracy, completeness, and citation integrity before the output reaches the user. Microsoft's chief marketing officer for AI at Work, Jared Spataro, described the workflow as "GPT drafts, Claude reviews... before it's delivered" and said the company expects it to run in both directions in the future, with Claude drafting and GPT critiquing.

The result, by Microsoft's own assessment, is measurably better output. On the DRACO benchmark — a 100-task deep-research evaluation across ten domains published by Perplexity researchers and academic co-authors in February 2026 — Researcher with Critique scored 57.4. That represents a 13.8 percent improvement over Perplexity Deep Research running Anthropic's Claude Opus 4.6 alone, which had been the previous top performer. These figures were self-assessed by Microsoft using GPT-5.2 as the evaluation judge; the methodology followed the original DRACO paper's protocol, but the results have not been independently replicated by a third party. Steve Gustavson, Microsoft's corporate vice president for design and research, told GeekWire that in Researcher, "Claude has proven to be a fantastic synthesizer and sort of check on what the GPT models might be doing," while also noting that Microsoft continuously evaluates the performance of single-model versus two-model configurations.

Critique ships alongside Council, which runs multiple models on the same prompt simultaneously so users can compare where GPT and Claude agree and diverge. CEO Satya Nadella highlighted Council publicly. Third-party coverage has reported higher compute costs for both features, though Microsoft has not published detailed pricing for the additional model calls involved.

The multi-model approach reflects a broader competitive reality: Microsoft's bet is that enterprise differentiation now sits at the context and orchestration layer — who controls how models are combined and governed — rather than at the raw model layer, where OpenAI, Anthropic, and Google continue to leapfrog each other. Microsoft, OpenAI's largest financial backer, now formally incorporates Anthropic's Claude into specific Copilot and Frontier features and markets Copilot as "model-diverse by design."

What Microsoft's Agent Stack Does Not Yet Prove

The three moves reinforce each other structurally: E7 supplies the deployment and oversight layer; computer-using agents expand what those agents can do inside software; and Critique embeds the premise that no single model is sufficient, so a second should check its work. Together they form the most coherent publicly available commercial stack for an autonomous-agent enterprise.

What Microsoft has not demonstrated is that any of this delivers durable value at the scale its own agent counts imply. Futurum Group analyst Keith Kirkpatrick raised the central unresolved question: how many organizations running Copilot Cowork in the Frontier program will actually move it to production by the end of 2026, and what will stop those that don't? Computer-using agents touch live systems. Agent 365 can govern access and audit trails. Critique may reduce specific categories of research error. None of these eliminates liability for bad agent actions, data exposure, or workflow failures — and the adoption figures Microsoft cites, including the tens of millions of agents in the registry and the 20 million paid Copilot seats reported at the April 29, 2026 earnings call, all originate with Microsoft.

The honest assessment is that Microsoft has assembled an unusually coherent commercial stack for the agentic enterprise. Whether enterprises will grant autonomous agents enough authority to justify the stack's cost is the question that production deployments, not preview counts, will answer.

Frequently Asked Questions

What is Microsoft 365 E7 and what does it cost?

Microsoft 365 E7, also called the Frontier Suite, became generally available on May 1, 2026, at $99 per user per month with Microsoft Teams, or $90.45 without it. It bundles Microsoft 365 E5, Microsoft 365 Copilot, the Entra Suite, and Agent 365 — the governance control plane for managing autonomous AI agents — into a single subscription tier.

How do computer-using agents in Copilot Studio work?

Computer-using agents interact with web and desktop applications through the screen — clicking, typing, and reading the interface — rather than through a programmatic interface. This allows enterprises to automate workflows in legacy applications that have no integration layer, with OpenAI's computer-use architecture and Claude Sonnet 4.5 available as the production-grade models at launch.

What is the Critique feature in Copilot Researcher, and does it improve accuracy?

Critique is a workflow inside Microsoft 365 Copilot's Researcher agent in which OpenAI's GPT produces an initial response and Anthropic's Claude reviews it for accuracy, completeness, and citation quality before delivery. Microsoft's own evaluation on the DRACO deep-research benchmark showed a 13.8 percent improvement over the previous top-performing system, though those figures were self-assessed and have not been replicated independently.

What is Agent 365 governance and why does it matter for enterprise AI?

Agent 365 is the control plane Microsoft built to manage autonomous agents at scale, covering agent identity through Entra, security monitoring through Defender, and compliance logging through Purview. It includes a Shadow AI detection page that identifies and can block unauthorized agents — including third-party tools — giving IT administrators a single registry for every agent operating in their environment.