Sources checked:
What changed
Between 28 and 29 September, three vendors announced different layers of control for AI agents. OpenAI began rolling out Dots, always-on agents with cloud computers and connected apps, to eligible Pro and Business Premium users; its Enterprise beta is off by default and requires an administrator to enable it. Google Cloud added preview-level, resource-specific IAM controls for individual Gemini Enterprise apps and data stores. NVIDIA announced an Open Agent Safety Platform: it says the broadly available OpenShell runtime can trace actions and enforce policy, while its Sentry reference design can monitor and quarantine agents outside the application layer. These are vendor product descriptions and rollout claims. They do not independently prove that a particular deployment is safe, reliable or profitable.
Business implications · analysis
The commercial decision is moving beyond model selection. Before assigning an agent an ongoing responsibility, a business needs to define the agent’s identity, the data and tools it may reach, the difference between reading and changing records, which actions require approval, who reviews its audit trail and how it is stopped. Personal assistants and organization-owned agents should not share an undefined permission model. Start with a narrow workflow in a test environment, grant the minimum access needed, measure completed work and exceptions, and expand only after the approval and recovery paths have been exercised. In procurement, compare administrative controls, logging, identity isolation and shutdown options alongside model quality and price.
What to watch next
Watch actual regional and plan availability, pricing after included allowances, the reliability of approval prompts, integration-level permission detail, false alarms in monitoring and the time required to recover from a wrong action. One scenario is that stronger controls make more persistent workflows acceptable. Another is that supervision and exception handling remain expensive enough to limit deployment to narrow, high-value duties. Treat vendor demonstrations as starting hypotheses and require operating evidence from your own workflow.
Primary sources
Company statements are attributed to the publisher. Our analysis is separate; forward-looking points are questions to monitor, not promised outcomes.