Summary: At VAST Forward on February 25, 2026, VAST Data announced PolicyEngine and TuningEngine, both due to ship with its AI OS before the end of 2026. PolicyEngine enforces agent permissions inline at the data layer before actions run, while TuningEngine turns outcomes and human feedback into LoRA, SFT and RL candidate models. The same NVIDIA-bound CNode-X stack makes governance and self-learning infrastructure, but concentrates adoption around one hardware path.
February 25, 2026, Salt Lake City: at the VAST Forward 2026 user conference, VAST Data moved further away from its storage-vendor roots. The company announced two new compute components for its AI Operating System — PolicyEngine and TuningEngine — built so that agents inside enterprises do not just run fast; they stay governable and keep improving themselves. CEO Renen Hallak framed it plainly: the system the company set out to build at its founding “has now come to life.”
Both engines are due to ship with the VAST AI OS before the end of 2026. Between them they patch the two holes today’s agent systems most conspicuously leave open — governance before an action executes, and learning after it has run.
PolicyEngine: Vetting Agent Actions Before They Run
The first component is an inline policy enforcer embedded in the data layer. Any access an agent attempts — shared memory, external tools, knowledge bases, even other agents — sits under its authority. Fine-grained explicit permissions pair with AI-derived contextual conditions, and the verdict lands before execution rather than during a later review of the logs.
Zero trust runs through the surrounding design. Agent actions leave tamper-proof traces and logs — which is what makes the system “observable, explainable, auditable.” For enterprises moving agents into production, the effect is IAM-grade governance raised to the agent-behavior layer: permissions live in the data plane, not in a prompt.
TuningEngine: Feedback In, Better Models Out
The loop’s other half belongs to TuningEngine. It keeps capturing outcomes from agentic pipelines together with curated human feedback, feeds that material automatically through LoRA, SFT, and RL fine-tuning pipelines to produce candidate models, then evaluates and benchmarks those candidates inside the AI OS itself. The final step is a human sign-off or an automatic deployment.
Add the existing AgentEngine — the serverless agent runtime that coordinates multi-agent workflows, model invocation, and tool use — and the target comes into view: a “thinking machine,” in co-founder Jeff Denworth’s words, cycling through observe, reason, act, evaluate, improve, then starting over.
CNode-X: Compute, Bound Tight to NVIDIA
The same event produced CNode-X, co-designed with NVIDIA. AI OS software runs directly on NVIDIA GPU servers as certified configurations, with Cisco and Supermicro among the certified builders. On top of NVIDIA cuDF, VAST has open-sourced its Sirius query engine, claiming query time down by up to 44 percent and query cost down by up to 80 percent, with cuVS vector-search acceleration and NIM microservices built in.
For long-context multi-agent inference, the item that counts for more is support for NVIDIA Context Memory (CMX). Shared KV cache access is accelerated by pairing BlueField-4 DPUs with Spectrum-X Ethernet, cutting time-to-first-token in long-context workloads at the source. This software-hardware binding rides the demand engine behind NVIDIA’s record $68.1 billion quarter, reported the same day (see NVIDIA’s FY2026 Q4 earnings). Three DataEngine blueprints went open-source as well: video intelligence, enterprise document RAG, and genomics.
CrowdStrike, Polaris, and Cosmos
Security integration is the other throughline. CrowdStrike is being embedded into the AI OS and PolicyEngine to monitor data access, admin activity, and workload behavior; correlate the signals with endpoint, identity, and threat intelligence; and, when needed, trigger quarantine, access restriction, workload isolation, and preservation of forensic records. Notably, parts of CrowdStrike’s own infrastructure already run on VAST.
Management moves to the new Polaris control plane, a Kubernetes-based multi-tenant design with a lightweight agent per node that manages fleets spread across on-premises, neoclouds, and public clouds — included with the AI OS at no charge. The Cosmos partner program launched in step: Mirantis’s k0rdent as the inaugural partner, South Korea’s Lablup joining with Backend.AI for sovereign AI, and Firmus Technologies appearing as a neocloud customer.
Where This Leaves Platform Teams
Three observations. First, governance keeps sinking down the stack: agent permissions enforced at the data layer are harder to bypass than app-layer or prompt-level controls — a real architectural pull for compliance-heavy buyers. Second, closed-loop economics: TuningEngine pulls “feedback to fine-tune to evaluate to deploy” into the infrastructure, turning agent improvement from a project cadence into an operational one. Third, lock-in risk: CNode-X, cuDF, CMX, and NIM all stack on NVIDIA, so adopting this means choosing the chip and the software stack in one move. Between now and the end-of-2026 release, benchmarking open-source query routes and alternative agent-governance approaches is time well spent.