On October 29, CIQ introduced a technical preview of Stratum Kubernetes Engine (SKE). General availability is expected November 5. The free engine uses Apache 2.0; Stratum Pro is the commercial edition.

One package does not mean identical permissions

The announcement describes a signed artifact for installation, joining and upgrades. Its MCP server divides 37 tools into read, write and destructive tiers. Unenabled tiers are not registered. SIExpo has not audited this mechanism.

The product description specifies read-only access by default. It describes refusals to read Secrets or write RBAC rules, and distinguishes tools using a user's Kubernetes credentials from administrative operations using node credentials. Stratum Pro adds multi-cluster management and single sign-on. CNCF conformance is a development goal; FIPS and STIG are not offered at launch.

What should an agent actually be allowed to do?

Our editorial assessment is that an agent interface should be judged by its permitted actions. An illustrative scenario is an assistant diagnosing a stalled workload. Its first useful result would be to find relevant events, explain dependencies and propose a verifiable next step. Permission to modify resources is a separate operator decision, rather than a prerequisite for every diagnostic task.

For a trial deployment, we would begin with an isolated cluster and describe precisely which objects are visible. We would then test refusals: what does the agent see when permission is absent, and can text inside a log redirect its intent? This is a proposed evaluation approach, not behaviour we have established in SKE. A readable refusal is useful only when the same boundary is enforced at the actual operation.

Recovery would receive particular attention. Suggesting an upgrade, executing it and restoring a snapshot are different decisions with different consequences if something goes wrong. In our view, separate access tiers are useful when operators can assess those decisions independently and retain an auditable record of authorised actions. A convenient interface does not remove that responsibility.

What still needs practical verification

SIExpo has not installed SKE, tested recovery from a failed upgrade or established production certification. An evaluation should identify the exact version and distinguish the free engine from paid fleet management. The verified development is the technical preview with specifically described agent boundaries. November 5 remains an announced future milestone.

*AI illustration: a conceptual cluster and bounded agent area, not a real control panel.*

*Editorial verification: 30.10.2026, 09:13:25 +0200 (Europe/Sofia).*