AI agent governance moves from theory to production reality
AI agent governance is the set of policies, controls, and technical safeguards that define which autonomous agents can exist in an enterprise, what systems they may access, and how their actions are monitored, audited, and limited across cloud and on‑premises environments. As enterprises move beyond pilots, this discipline is shifting from a theoretical concern to day‑to‑day infrastructure. Analysts already warn that AI agents behave like non‑human employees: they hold credentials, reach into business systems, and can trigger irreversible workflows. According to Gartner, an average global Fortune 500 company may be running more than 150,000 agents by 2028, a scale that makes manual oversight impossible. At the same time, Okta reports that 90% of enterprise agents are over‑permissioned, with more than half touching sensitive data. The result is a race among major cloud vendors and automation platforms to embed enterprise AI controls before autonomous agent sprawl turns into a security incident.
Microsoft builds cloud PCs as fenced gardens for AI agents
Microsoft’s Windows 365 for Agents shows how cloud AI infrastructure is evolving to contain autonomous behavior instead of scattering it across desktops and servers. The service runs AI agents inside secure cloud PCs, where they can use browsers, desktop applications, file shares, and legacy systems while remaining inside a governed environment. Organizations define agent identities, access rights, and configurations through existing tools such as Microsoft Entra ID and Intune, extending familiar user controls to non‑human actors. Multi‑step workflows execute within defined security boundaries so agents cannot move laterally into unintended systems. Microsoft’s approach addresses a concern raised in a recent Cloud Security Alliance report: agent activity should have the same rigor and traceability as human users. By turning the cloud PC into a dedicated execution zone, Windows 365 for Agents aims to reduce autonomous agent security risks like data exposure and uncontrolled access to fragile production systems.
Okta and AWS target the identity crisis in autonomous agents
Okta is positioning identity as the backbone of AI agent governance, extending Okta for AI Agents to cover new ecosystems and any identity provider. A new integration with Amazon Bedrock AgentCore adds identity lifecycle management for AI agents built on AWS, helping security teams discover, onboard, and govern agents across cloud AI infrastructure without replacing existing stacks. Ely Kahn, Okta’s Chief Product Officer, argues that security and IT leaders need clear answers to basic questions: where agents run, which systems they connect to, and what they are allowed to do. Okta cites a sharp mismatch between deployment and control, noting that 90% of enterprise agents are over‑permissioned and more than half access sensitive information. By tying agent accounts to central policies and enforcing least privilege across platforms like Salesforce Agentforce, ServiceNow, and Google Vertex AI, Okta is trying to make enterprise AI controls scale at the same pace as agent development.
Automation Anywhere’s EnterpriseClaw wraps powerful agents in control layers
Automation Anywhere’s EnterpriseClaw highlights why governance must keep up with more capable autonomous agents. Inspired by Nvidia’s OpenShell runtime, the company defines “claw‑style” agents as those with device‑level file access, dynamic tool creation at runtime, and direct interaction with the computer screen. On their own, such agents can “access pretty much everything,” a serious problem for banks, healthcare systems, or factories with strict boundaries. EnterpriseClaw keeps those capabilities but surrounds them with centralized governance, credential controls, and observability so agents run close to where data lives, including behind firewalls or in environments that never reach public cloud. Partnerships reinforce this control stack: Cisco for infrastructure security, Nvidia for OpenShell, Okta for identity management, and OpenAI to supply models such as GPT 5.5. The message is clear: as agents become more human‑like in what they can do on endpoints, governance infrastructure must become more systematic and automated.
Alibaba Cloud turns Qwen into a full agentic stack for enterprises
Alibaba Cloud’s Qwen roadmap shows how model providers are expanding into AI‑native clouds so enterprises can deploy agents with built‑in guardrails. Its Qwen3.7‑Max model, ranked fifth globally and first among Chinese models in Artificial Analysis’s Intelligence Index, now sits at the core of Qwen Cloud, an AI‑native platform for applications and agents. Qwen Cloud exposes three entry points: a Skills portal for agents, a command‑line interface for workflow integration, and a website for human users. The Skills layer converts common cloud operations into controlled capabilities that agents can call, while the platform combines Qwen models with open‑source and third‑party options across text, vision, audio, image, video, and embeddings. Training initiatives for enterprises and students aim to close the skills gap so teams can configure and supervise agents effectively. By tying models, tools, and training into one environment, Alibaba Cloud is treating AI agent governance as part of mainstream cloud operations rather than a separate add‑on.