Nutanix NAI 2.8: governing agentic AI across hybrid cloud
Nutanix has made Enterprise AI 2.8 generally available, with centralized control for inference and agents, a production-ready MCP gateway, and private inference capabilities. For Belgian and French companies, the value is less about another product than an architecture pattern: bringing AI closer to applications and data without giving up governance.
1. A control layer for agents
According to Nutanix, NAI 2.8 adds an MCP gateway to govern how agents connect to applications and data. The offering fits an architecture spanning virtual machines and containers, while Nutanix Kubernetes Platform 2.19, announced as coming soon, is expected to add an AI catalog for agentic applications.
The distinction between current availability and roadmap matters: NAI 2.8 is available, while NKP 2.19 is forthcoming. Companies should qualify what is actually delivered, its limits, support, and observability before building a critical service.
2. What this changes for a Belgian or French company
For an SME or mid-market company, a hybrid platform can reduce the need for separate stacks for models, legacy applications, and new agents. For a large enterprise or public administration, it can help place selected workloads close to sensitive data, segment environments, and centralize policy. In both cases, value depends on proving who called which tool, with what data, and what result.
An agent connected to Odoo Enterprise illustrates the issue: reading a product record, preparing a quote, and changing an order do not carry the same risk. An MCP gateway therefore needs technical identities, tool allowlists, human approvals, action limits, and logs that correlate with the ERP. Private inference can reduce some data transfers; it does not automatically make the entire system sovereign or compliant.
3. Underside analysis: hybrid is a policy, not a location
Hybrid cloud becomes useful when it implements business classification: which data may leave, which models may be called, which actions are authorized, and what continuity is required. Installing a model locally without governing connectors, secrets, or updates moves risk rather than reducing it. Conversely, a cloud service can suit a non-sensitive workload when contracts, residency, logging, and reversibility are controlled.
For RAG, agents, and Apple Enterprise workflows, architecture should separate indexing, inference, and action. Documents can remain in a private perimeter, models can be selected by use case, and any write into Odoo or another business application can pass through dedicated approval. This separation also makes cybersecurity testing and model replacement easier.
4. Operational recommendation
Before selecting an agent gateway, test a complete scenario: agent identity, MCP discovery, data reading, refusal of a prohibited tool, human approval, audit log, model outage, and emergency revocation. Then measure latency, cost, manual recovery, and outcome quality—not model performance alone.
Practical priority: establish a “data × tools × actions × environments” matrix for each agent, then verify that the platform actually enforces and logs those rules.
Frame an agentic architecture