Google has begun shipping the governance primitives it promised in April with the Gemini Enterprise Agent Platform—the official successor to Vertex AI—while OpenAI flipped the switch on usage-based billing for agent features in its Windows desktop app on July 6. The twin moves mark a sharp pivot from free experimentation to a world where autonomous agents carry real operational costs and security enforcement lives in the infrastructure, not just the application layer.

The Agent Identity Lockdown

On April 22 at Cloud Next '26, Google announced that future Vertex AI roadmap updates would come solely through the Gemini Enterprise Agent Platform. But the most consequential additions—Agent Identity, Agent Registry, and Agent Gateway—did not land until later, shifting the platform from a model hub into a control plane for machines acting on corporate data at machine speed.

Agent Identity assigns every agent a unique cryptographic identifier. Every API call, file access, or database query is signed, logged, and traceable to that identity, just as Entra ID or Active Directory would track a human employee. Agent Gateway sits in front of every tool and data source an agent touches, enforcing least-privilege policies and filtering prompts through Google’s Model Armor to block prompt injection and data leakage. Agent Registry provides a centralized catalog of approved agents, tools, and MCP servers—a bulwark against the shadow AI that IT departments dread.

The platform supports more than 200 models via Model Garden, from Anthropic’s Claude to Meta’s Llama, alongside Google’s own Gemini, Gemma, and Lyria families. The Agent Development Kit (ADK) 2.0 models orchestration as a graph, with native OpenTelemetry export and support for Python, TypeScript, Go, Java, and Kotlin. Agent Runtime claims cold starts under one second and multi-day persistent workflows.

Critically, the entire stack holds ISO 42001 certification for AI management systems. “The original Vertex AI was designed for the complexity of the early generative AI era,” Google Cloud VP Michael Gerstenhaber said in the launch. “Today, we’re managing agents interacting across multiple systems—and often without security and governance guardrails.”

OpenAI Fires Up the Usage Meter

While Google hardened its governance stack, OpenAI began applying a different kind of guardrail: cost. On July 6, ChatGPT Workspace Agents and the Excel/Sheets integration for Enterprise and Edu customers moved from free preview to credit-based pricing. Input tokens, cached input tokens, and output tokens now each draw from a usage pool, with a typical GPT-5.5 run landing between 5 and 25 credits, according to OpenAI’s rate card.

Administrators who had wired agent workflows into daily operations suddenly faced a variable expense line. Then on July 9, OpenAI launched ChatGPT Work, a new desktop app for Windows that combines Chat, Work, and Codex into a single client. Work can access local files and desktop apps with user permission, while the PowerPoint integration went GA for Enterprise users, deploying through Microsoft 365 with role-based controls where supported.

For Windows-heavy shops, the math flipped overnight: features like “summarize this Excel sheet” or “generate this PowerPoint deck” stopped being perks of a flat subscription and became metered consumption that demands FinOps discipline and clear usage limits.

What the Governance Race Means for Windows Shops

Organizations running Microsoft 365 and Entra ID face a clear fork. On one side, Azure AI Foundry extends existing identity, compliance, and DLP policies to agents by default—a day-one governance win without new infrastructure. On the other, Google’s infrastructure-level Agent Identity creates a cryptographic audit trail that cuts across any model or tool, whether the agent lives on Google Cloud, on-premises, or in another provider.

For Windows admins, the decision hinges on where sensitive data sits and what kind of audit evidence regulators will demand. If agents touch customer financial records or healthcare data, infrastructure-layer enforcement becomes a compliance forcing function. If agents only augment productivity—drafting emails in Outlook, summarizing Teams chats—the native Entra ID inheritance inside Azure AI Foundry may be enough.

One concrete example: a Windows admin who allowed a ChatGPT Work agent to scrub a large Excel file on July 5 paid zero extra. The same task on July 7 consumed credits from a pooled budget, visible in OpenAI’s admin analytics. That shift puts agents squarely in the same category as cloud VMs or API calls—resources that need forecasting, capping, and departmental chargebacks.

Anthropic added its own variable on July 7, expanding Claude Cowork from desktop to web and mobile. Its enterprise tier at $100/user/month emphasizes task execution across Google Workspace, Salesforce, and Docusign, but governance and observability remain largely delegated to third parties, limiting its appeal for compliance-driven IT functions.

How We Got Here: A Timeline of Enterprise Agent Chaos

  • 2025: Gartner’s Hype Cycle places agentic AI at the Peak of Inflated Expectations, projecting 40% of enterprise apps will embed task-specific agents by year-end 2026.
  • Early 2026: An OutSystems survey of 1,900 IT leaders finds 96% of enterprises already run agents in production, yet only 12% claim they can govern them.
  • April 22, 2026: Google launches Gemini Enterprise Agent Platform at Cloud Next, retiring the Vertex AI brand.
  • July 6, 2026: OpenAI introduces credit-based billing for Workspace Agents; Excel/Sheets and PowerPoint integrations leave free tier.
  • July 7, 2026: Anthropic expands Claude Cowork to the web and mobile.
  • July 9, 2026: ChatGPT Work desktop app for Windows arrives, bundling agentic capabilities with local file access.
  • July 17, 2026: Google announces deprecation of Grok 4.1 model family on its platform, with a shutdown deadline of August 20, forcing affected teams to migrate.

The governance gap—96% deploying, 12% governing—is closing faster than budgets have anticipated. Gartner projects that 40% of enterprise applications will embed task-specific AI agents by the end of 2026, up from under 5% in 2025.

What to Do Now: A Practical Checklist for Windows Admins

Inventory your agent surface area

Which departments use ChatGPT for Excel, PowerPoint, or workspace tasks? Which internal tools connect to third-party models? Map every agent that touches company data, even indirectly.

Set hard limits in OpenAI’s console

Log into the admin portal for your Enterprise or Edu tenant and apply credit caps per user and per workspace. Enable usage alerts before a weekend spike burns through a month’s budget.

Evaluate identity inheritance vs. cryptographic identity

  • If your organization is deeply Microsoft 365/Entra ID native: test Azure AI Foundry’s agent capabilities alongside your existing DLP and conditional access policies. The per-seat pricing (approx. $30 plus $15 for Agent 365) offers predictable costs.
  • If you use multi-cloud or need auditable proof of agent actions: pilot Google’s Gemini Enterprise Agent Platform. Its Agent Identity and Agent Gateway provide a single policy enforcement point across models and cloud environments, with ISO 42001 certification to back up claims.

Pressure test your agent’s least-privilege stance

Do not let agents inherit a user’s broad permissions. Use Agent Gateway (Google) or Azure policies to scope permissions to the minimum data required. Test with a read-only agent first, even if the goal is a read-write workflow.

Track model deprecations

If you call Grok 4.1 models through Google’s platform, migrate to 4.2 or 4.3 before August 20. Multi-model flexibility cuts both ways: it decouples you from a single vendor, but increases operational overhead.

The Road Ahead

The competition is no longer about which model scores highest on a benchmark. It is about which platform can prove—in an audit, on a budget sheet, and during a security incident—that every action an agent took was authorized and traceable. For Windows admins, that means the agent conversation must move from an engineering curiosity to a standing infrastructure item, governed by the same rigor applied to user accounts, VMs, and API keys.

Watch for Microsoft’s response: tighter integration between Azure AI Foundry and Entra ID’s identity-governance features could close the gap quickly. Also monitor the Grok 4.1 sunset on August 20 as a real-world test of multi-model platform reliability. And expect the next OWASP agentic security report to further sharpen the requirements for infrastructure-layer enforcement. The window for uncontrolled agent experimentation is closing—not because of policy, but because the meters are running and the auditors are coming.