Mirantis at the KubeCon + CloudNativeCon North America 2026 conference made available a consulting service through which IT teams can train artificial intelligence (AI) agents in either an OpenStack or Kubernetes environment using a set of best practices that make them easier to deploy and manage.

The Mirantis Agent Platform comes with a set of built-in Prism agents and third-party agents to proactively detect production issues, diagnose root causes and propose fixes. IT teams can also work with forward deployed engineers (FDEs) that Mirantis has previously made available to build additional AI agents to automate additional IT operational workflows. The platform also integrates with existing developer tools to enable IT teams to manage any AI agent in any IT environment.

Ronen Kofman, vice president of product management for enterprise at Mirantis, said the overall goal is to make it possible to safely deploy AI agents that will not, for example, inadvertently delete a production database. IT teams are able to define what each agent can access and then deploy them in an isolated sandbox to limit unauthorized access or other rogue behaviors that, if they do occur, can be immediately terminated. IT teams can also require a person to approve proposed fixes before agents apply them and set spending limits at the organization, project and individual agent levels. Additionally, audit trails record every agent action to support operational reviews, incident investigations and compliance.

In general, IT teams are struggling to inject AI agents that generate probabilistic outputs into IT workflows that are largely deterministic in the sense they need to be executed the same way every time. AI agents that are not properly constructed will also require constant supervision, noted Kofman. Too many IT teams are deploying AI agents in sandboxes without realizing that a sandbox is only one piece of a larger system that needs to be deployed, he added.

AI agents deployed by different members of an IT team are also likely to come into conflict with one another. In effect, IT teams that build their own AI agents without an appreciation for the best practices and sandboxes needed to manage and secure them are creating a recipe for disaster, said Kofman.

Many of those IT teams are also in danger of becoming overly dependent on a small number of engineers that are the only ones on the team that know how to maintain those AI agents, he noted. If those engineers for some reason wind up no longer working for the organization, then the investments made in AI agents are for all intents and purposes lost, said Kofman.

It’s not clear how many AI agents that IT teams have deployed themselves but most are clearly experimenting. It’s becoming apparent that IT teams will not be able to manage applications at scale in the AI era without relying more on AI agents. The challenge is acquiring and retaining the expertise required to successfully deploy what might easily become thousands of AI agents.