AI agents are leaving the experimental phase. They write code, launch jobs, access repositories and clusters – and thus behave like additional employees without a personnel file. Two current reports show what currently determines the productive use of AI agents: governance in operations and the question of which data flows outwards in the process.

KubeCon 2026: Governance instead of pilot project

At KubeCon North America 2026, one focus was on controlling AI agents in Kubernetes environments. Kubernetes is the widespread platform for operating containers – isolated runtime packages for applications – automatically. This is precisely where agents that trigger actions independently are increasingly running.

Three topics were particularly highlighted:

  • Access rules, presented by Mirantis: Which resources may an agent see, which may it modify?
  • Cost control with Splunk: Agent runs generate computing load and token consumption that must be visible and attributable.
  • Verifiable identities via ANS: An agent should be verifiably who it claims to be.

Together, these three points describe the level of maturity that companies must reach when using agents. An agent without a unique identity cannot be assigned to any action after the fact. An agent without access rules works with the rights of whoever started it – in doubt, with too many. And an agent without cost transparency only becomes noticeable when the bill arrives.

Identity as the foundation for everything else

Technically, verifiable identities are the prerequisite for the other two building blocks. Only once it is clear which agent is making a request can permissions be finely graded and costs allocated to the party responsible. In classic environments, user accounts and service accounts handle this; with agents, there is the added factor that they are created dynamically, call each other and act on behalf of people.

In practice, this means: agents belong in the same identity and permission management as other system components – not in a parallel world of API keys in configuration files.

Costs become a management tool

The fact that cost control gets its own slot at an infrastructure conference is remarkable. With classic applications, the cost per request is largely predictable. Agents, by contrast, decide for themselves how many steps they take, how often they query a model and which tools they call. The effort for the same task can vary considerably depending on the solution path.

Observability – that is, the systematic collection of metrics, logs and traces – thus becomes a business management topic rather than an operational one. Anyone who does not measure which teams and which use cases cause which costs can neither plan budgets nor decide where agents actually pay off.

Windows 11 collects data on coding agents

The second report concerns the developer workstations themselves. According to the report, Windows 11 collects usage data from coding agents and developer tools and transmits it to Microsoft together with the optional diagnostic data. Optional means: the extended diagnostic data transmission can be influenced in the system settings – so it is not an unavoidable component.

In terms of content, this is about telemetry, i.e. usage and status data, not the content being processed. Nevertheless, the point is relevant for companies: information about which tools are used and to what extent is part of the internal technology picture. Where works agreements, data protection impact assessments or customer contracts impose strict requirements, a deliberate decision is worthwhile instead of a default setting.

What the two developments have in common

At first glance, cluster operations and operating system telemetry stand side by side. In fact, they describe two sides of the same question: who has insight into what agents do – and who retains control over it?

In the data centre, the aim is to equip agents with identity, permissions and a cost budget. On the workstation, the aim is to know which usage data leaves one's own area of responsibility. Both are governance tasks that traditionally sit in different departments and must be brought together in agent projects.

Practical starting points

  1. Inventory: Record which agents and AI-powered developer tools are already in use – often more than officially known.
  2. Assign identities: Each agent receives its own, traceable account instead of shared credentials.
  3. Minimise permissions: Access only to the systems required for the respective task.
  4. Make costs visible: Evaluate consumption per team and use case before scaling the deployment.
  5. Review telemetry settings: Configure and document the data transmission of the workstation systems deliberately.

For web and software projects, this means: agents can be used productively, but not without the same care that already applies to access, logging and budgets. Anyone who considers identity, permissions and cost measurement from the outset avoids costly retrofits later during ongoing operations.

And anyone who knows the telemetry of their own development environment can provide reliable information to customers and data protection officers instead of having to research it after the fact.

Sources