OpenShift 5 brings AI incident investigation with human approval for fixes

Red Hat released OpenShift 5 on November 9, adding AI-led incident investigation and update risk assessments to its enterprise application platform. The release is available through the customer portal to existing subscribers running containers, virtual machines and AI workloads.

The change brings investigation and proposed remedies into the management console. In supervised mode, the system presents its findings and next steps in natural language, then waits for an operator’s confirmation. Red Hat describes a workflow that helps teams correlate operational evidence before deciding what to change in a cluster.

An autonomous mode, offered as a technology preview, lets the agent investigate more deeply on its own. It still requires final human approval before executing changes. That makes the release status of the platform and the maturity of its autonomous investigation feature separate considerations for customers. Red Hat’s official demonstration description, published November 2, shows an agent querying workloads, logs and resource metrics after an alert, then presenting a remediation plan. Operators configure approval policies and inspect resolution results and audit records in the investigation history. This is the company’s illustrated workflow.

Operators can also ask the assistant to build metrics dashboards. A separate official demonstration describes generating CPU and memory utilization charts and adding them to the console. It also shows querying active alerts, inspecting platform responses and switching into troubleshooting to review a proposed remedy.

The MCP server and gateway for controlled agent access also reach general availability. According to Red Hat’s developer overview, the server uses existing identity systems and Kubernetes role-based permissions, defaults to read-only access, and tags actions for auditing. Example tasks include inspecting pipeline runs and monitoring virtual-machine resources. The gateway screens requests and responses before they reach MCP servers. Red Hat says its checks cover personal information, prompt injection and unsafe content. A separate Agent Sandbox for executing untrusted agent code in virtual-machine isolation is being offered as a technology preview.

Update paths and support terms

AI-generated update assessments are in technology preview. The console checks available versions and cluster compatibility, including optional components, then produces a report for administrator review.

The assessment requires administrative privileges, installed AI services and a language model configured by the customer. That provider processes the cluster-state data. The agent checks cluster health, nodes, add-on compatibility and deprecated interfaces inside a read-only sandbox. It produces a risk assessment and recommendation; an administrator separately starts the actual update.

Technology previews fall outside Red Hat’s production service-level agreements, and the company advises against production use.

Customers can update in place from 4.22 while retaining CoreOS 9. New 5.0 installations default to CoreOS 10.2. The release notes say that a subsequent move to 10.2 must cover all nodes; mixing 9-series and 10.2 nodes within a cluster is unsupported.

The release notes also flag approximately 30–50% lower throughput for messages smaller than 4KB on the RHEL 10 stream versus RHEL 9. The regression is most pronounced at sizes such as 64 bytes and 1KB and diminishes at 4KB or above.

Starting with 5.0, the standard lifecycle grows from 18 to 24 months: one year of full support followed by one year of maintenance support. Red Hat is also aligning support for selected platform add-ons, with their own documentation defining compatibility and coverage.

Beginning with 5.2, every third release is scheduled to qualify for Extended Life Cycle support. An add-on of up to three additional years would bring those designated releases to five years in total. Red Hat plans to ship them between June and September each year.